ProgrammableMarketing/usesalus.ai (opens in a new tab)
PDF

Views

  • Plan
  • Roadmap
  • Research

On this page

  • The brief
  • Start here
  • Idea portfolio
  • Effort vs impact
  • This week · 5
  • This quarter · 13
  • Big bets · 3
  • Wildcards · 3
  • How we can help
Generated Nov 10, 2026 1 view
  1. Programmable
  2. Playbooks
  3. Salus
Salus logo

growth playbook · Nov 10, 2026

usesalus.ai (opens in a new tab)

Salus growth playbook

Diagnosis

Salus is a two-person YC W26 company with a sharp product story ("Agents don't crash. They commit."), clear public pricing and a founding BDR and founding engineer being hired now. The funnel around that story is thin, though. The docs that every page points to show up as an empty page titled "Salus Cloud" to any reader that doesn't run JavaScript.

Every "Get started" button, even on the $500 cancel-anytime plan, leads to an application that doesn't mention a price. The headline statistics have no sources. And the 10-page site has no integration, use-case, comparison or customer pages. This playbook starts with readable docs, a clear pilot path and sourced numbers.

Next it gives the new BDR something to sell: a free shadow replay of a prospect's own agent actions, receipts-based posts and trigger-based lists. Then it builds the developer funnel (integration and use-case pages, a shadow-only starter plan, a Show HN) and the proof that enterprise buyers ask for (a security pack, latency numbers, a public benchmark).

Growth Score

6/10

Biggest Opportunity

Let developers and their coding assistants find, read and try Salus on their own, then use their own agent logs as the sales demo.

Biggest Risk

Arcade, Permit.io and Arcjet are already chasing the same "control every agent action" space with self-serve signups, and one reviewer called the core proxy replicable in months. A founder-only, application-gated funnel may be too slow to win developers before the category settles.

01

Start here

The three moves to make first. Each one is written up in full below.

  1. 1

    Make the docs readable without JavaScript, so developers, search engines and coding assistants can find the integration

    Open the quickstart with JavaScript disabled, confirm it is empty, and switch the docs build to static or server-rendered output.

    High impact First result in 2-4 weeks
  2. 2

    Show the price and the next step on /pilot, so "Get started" doesn't feel like a job application

    Add the plan parameter to the three pricing buttons and show it on /pilot.

    High impact First result in 2 weeks
  3. 3

    Source the 53% and publish the method behind the benchmark numbers

    Find the original source of the 53% statistic.

    Medium impact First result in 2 weeks

02

Your idea portfolio

24 ideas, grouped by when to run them. Open any idea for how to run it: the steps, who to approach, copy you can use, and how to measure it.

  • 8 ideas are marked Unexpected: plays you are unlikely to have tried.

    1 of these are things Programmable can build and run for you.

Every idea placed by the effort it takes against the impact it could have. Quick wins sit top left; select a number to open that idea. The filters below narrow the list by status, timing, category and what Programmable can build.

Effort vs impactEach number is an idea. Select one to open it.
Impact →
Quick wins · 2Big bets · 11Fill-ins · 8Reconsider · 3
Effort →

Showing 24 of 24 ideas

This week 5

Quick wins with a clear first action.

  • 01

    Salus sells "a pip install plus a few lines of code", and the homepage, product and integrations pages all send developers to docs.usesalus.ai. But the quickstart, integration and webhook pages each return only the title "Salus Cloud" to any reader that does not run JavaScript, and the docs have no sitemap.xml, robots.txt or llms.txt. Many link previews, lighter crawlers and AI coding and research assistants read pages without running JavaScript. So when a developer pastes the quickstart link into their assistant and asks it to add Salus to a LangGraph or Vapi agent, the assistant gets an empty page. In search, every docs page has the same title.

    High impactLow effort First result in 2-4 weeks Founder & CEO

    First action

    Open the quickstart with JavaScript disabled, confirm it is empty, and switch the docs build to static or server-rendered output.

    How to measure it

    Indexed docs pages with unique titles

    Now
    –
    Target
    Every docs page
    By
    30 days

    Steps

    1. 1

      Confirm the gap

      Fetch the quickstart, integration and webhook pages with JavaScript off and list what is missing.

      Founder & CEO 1 hour

    2. 2

      Pre-render

      Static export or server rendering for every docs route.

      Founder & CEO 2 days

    3. 3

      Title each page

      A unique title and description per page instead of "Salus Cloud".

      Founder & CEO 1 day

    4. 4

      Add sitemap and robots

      docs sitemap.xml and robots.txt, submitted in Search Console.

      Head of Growth 2 hours

    5. 5

      Publish llms.txt

      An index plus one full-text file with the quickstart, protect, webhook and executor patterns and framework examples.

      Founder & CEO 1 day

    6. 6

      Assistant test

      Ask three coding assistants to add Salus to a LangGraph agent from the docs URL; fix whatever they get wrong.

      Founding SWE / MLE 1 day

    7. 7

      Measure

      Indexed pages, docs search impressions and docs-to-/pilot clicks.

      Head of Growth monthly

    First result
    2-4 weeks
    Cost
    A few days of engineering time; no new tools
    Owner
    Founder & CEO

    Smallest test

    Pre-render the quickstart alone and check it with JavaScript off the same day.

    Why it works

    Pre-render or statically export the docs so each page ships its text in the HTML. Give every page its own title and description, and add a docs sitemap and robots.txt. Publish an llms.txt index plus one full-text file covering the quickstart, the salus.protect pattern, the webhook pattern (Vapi, Retell, HTTP), the executor pattern and each framework example. Then test it the way developers actually work: fetch each page with JavaScript off, and ask two or three coding assistants to integrate Salus starting from the docs URL alone.

    For a developer tool, the docs are the product page for the person who does the install. When the docs are readable, each topic can rank on its own for framework-specific searches, and assistants write correct integration code instead of guessing. That shortens the path from first visit to first protected route, and it costs a build setting, not a rewrite.

    Ready to use

    ChecklistReadable docs launch list

    [ ] Quickstart, integration and webhook pages show full text with JavaScript off
    [ ] Unique <title> and meta description per page
    [ ] docs.usesalus.ai/sitemap.xml and robots.txt
    [ ] llms.txt index and llms-full.txt
    [ ] Install command and salus.protect example in the first screen of the quickstart
    [ ] Search Console property for the docs subdomain
    [ ] Three coding assistants produce a working integration from the docs URL

    Risks and how to handle them

    • The docs framework can't export static pages

      Serve a pre-rendered HTML snapshot per route to non-JavaScript readers.

    • Public docs expose more than intended

      Publish only the integration patterns the marketing pages already describe; keep account-specific content behind login.

    Evidence

    • Your websiteDocs (fetched twice on November 10, 2026): docs.usesalus.ai, /docs/cloud-quickstart, /docs/integration and /docs/cloud-custom each return only the page title "Salus Cloud" to a reader that does not run JavaScript; docs.usesalus.ai has no sitemap.xml, robots.txt or llms.txt (all 404). The homepage, product and integrations pages send developers to these docs ("Read the quickstart", "Python integration", "Webhook integration").Source
    • Your websiteHomepage (checked November 10, 2026): title "Salus — Runtime authorization for AI agents", H1 "Agents that act, safely."; "53% of organizations have already had an agent exceed its permissions." with no source; "Measured results" of −70.6% policy failures and +40.4% compliant completions on τ²-bench and −20.5% agent-model spend per successful task on CarBench, with no link to a method, paper or repo; a Python sample ("from salus import Salus" … "salus.protect(...)" around a do_refund tool) with no install command, package or GitHub link; a 59-second airline-booking demo; no customer logos or testimonials.Source
    • Your websiteIntegrations page: frameworks LangChain, LangGraph, CrewAI, AutoGen, PydanticAI, LlamaIndex and Google ADK; voice and realtime Vapi, Retell, LiveKit Agents and Pipecat; MCP, HTTP, computer use and custom tool loops; three patterns (SDK, webhook for Vapi, Retell and HTTP, executor). None has its own page; all links go to the docs.Source
    • News coverageY Combinator: Salus is in the Winter 2026 batch with a team of 2 and the description "Guardrails to validate your agent's actions before they execute". Its launch post claimed 58% of blocked actions recovered and completed correctly, "up to 60% lower cost" on τ²-bench and a 52% average misalignment reduction on ODCV-Bench across 12 frontier models, with onboarding as a pip install plus a few lines of code.Source
  • 02

    The pricing page presents Pilot as a $500 a month, month-to-month, cancel-anytime plan. But every "Get started" button, including Partner and Enterprise, leads to "Apply for a founder-led pilot", which states no price, no plan and no length. The word "pilot" means both a paid plan and an application, and a visitor who picked Partner lands on the same four-field form. Meanwhile Arcade and Permit.io let developers start right away.

    High impactLow effort First result in 2 weeks Head of Growth

    First action

    Add the plan parameter to the three pricing buttons and show it on /pilot.

    How to measure it

    Pricing page to /pilot submissions

    Now
    –
    Target
    +25%
    By
    60 days

    Steps

    1. 1

      Pass the plan

      ?plan= on each pricing button.

      Founder & CEO 1 hour

    2. 2

      Show the plan

      Plan name, price, limits and terms at the top of /pilot.

      Head of Growth 2 hours

    3. 3

      Rename the form

      "Start your first protected route"; keep "Pilot" for the plan only.

      Head of Growth 1 hour

    4. 4

      Add two fields

      Framework or platform, and the first action to protect.

      Founder & CEO 2 hours

    5. 5

      State the timeline

      Reply in 12 hours, mapping call, shadow mode in week one.

      Head of Growth 1 hour

    6. 6

      Track

      Pricing views, submissions by plan, time to first call.

      Head of Growth weekly

    First result
    2 weeks
    Cost
    $0
    Owner
    Head of Growth

    Smallest test

    Show the selected plan and price on /pilot for two weeks and compare submissions.

    Why it works

    Pass the chosen plan to /pilot (for example ?plan=partner) and show it at the top with its price, limits and "cancel anytime". Rename the form to "Start your first protected route" so it reads as onboarding, not a gate. Add two optional fields: framework or platform (from the integrations list) and the action to protect first (refund, booking, write). Then spell out what happens next: a reply within 12 hours, a 30-minute route-mapping call, and shadow mode in week one.

    Buyers who have just read a public price expect to buy at that price. When the next step confirms the plan and shows a timeline, an application stops feeling like an unknown wait, and the two new fields give the founder a better first call.

    Risks and how to handle them

    • Enterprise buyers dislike seeing a form

      Offer the 30-minute demo link as an equal option for Enterprise.

    • Extra fields lower submissions

      Keep both optional.

    Evidence

    • Your websitePricing: Pilot "$500 / month", "Month-to-month · Cancel anytime", 10 agents and 200K action checks a month; Partner "$2,500 / month", 50 agents, 1.5M checks, priority support; Enterprise "Custom", unlimited agents, deployment review; "Unlimited seats"; shadow or enforcement per route; Incident Rewind included; checks stop at the ceiling with notice ("Enforcement never silently degrades"); no free tier or trial; every plan's "Get started" goes to /pilot.Source
    • Your websitePilot page: "Apply for a founder-led pilot." with a reply promised "within 12 hours"; a four-field form (Name, Work email, Organization, "What does your agent do?"); steps "Map the consequential action route", "Connect your existing agent and tools", "Run the route in shadow or enforcement"; no price, plan choice or pilot length on the page; a "Book a demo" link to a 30-minute calendar slot; "No mailing list."Source
    • Competitor researchArcade's homepage: "Govern every action for every AI agent", free to start with "Try Free" self-serve signup; Permit.io's homepage: "Runtime Access Control for the AI Era" with an MCP Gateway and a "Start building" signup; RuntimeWire reports Arcjet launched agent runtime security on September 17, 2026, with deterministic checks in application code that block tool, API and database calls before execution.Source
  • 03

    The homepage opens its problem section with "53% of organizations have already had an agent exceed its permissions" and gives no source. The "Measured results" (−70.6%, +40.4%, −20.5%) link to no method, paper or repo, and the launch post used different figures (58% recovery, up to 60% lower cost, 52% on ODCV-Bench). One public review already notes that the method is not explained. That hurts a company whose promise is agents that "can prove what they're about to do".

    Medium impactLow effort First result in 2 weeks Founder & CEO

    First action

    Find the original source of the 53% statistic.

    How to measure it

    Homepage to benchmarks click-through

    Now
    –
    Target
    10% of homepage visitors
    By
    60 days

    Steps

    1. 1

      Find the source

      Original study and wording for the 53%.

      Head of Growth 2 hours

    2. 2

      Footnote or replace

      Link the source or swap in an owned number.

      Head of Growth 1 hour

    3. 3

      Write the method

      Setup, models, policies, runs and baseline for τ²-bench, CarBench and ODCV-Bench.

      Founder & CEO 2 days

    4. 4

      Reconcile figures

      Explain why launch and homepage numbers differ.

      Founder & CEO 2 hours

    5. 5

      Publish /benchmarks

      Link each homepage number to it.

      Founder & CEO 1 day

    6. 6

      Share

      Post the method from the founder's account.

      Founding GTM / BDR 1 day

    First result
    2 weeks
    Cost
    $0
    Owner
    Founder & CEO

    Smallest test

    Add a source link under the 53% today.

    Why it works

    Footnote the 53% with its source, or replace it with a number Salus can stand behind. Publish one /benchmarks page that covers each benchmark's setup, models, policies, number of runs, baseline and result, and that reconciles the launch figures with today's homepage figures. Link it from every number on the homepage, and add a downloadable run log or repo where possible.

    Technical buyers check numbers before they repeat them to a security team. A sourced figure gets quoted, and an unsourced one gets questioned on the first call.

    Risks and how to handle them

    • The method invites criticism

      Publish limits openly; it reads as confidence.

    Evidence

    • Your websiteHomepage (checked November 10, 2026): title "Salus — Runtime authorization for AI agents", H1 "Agents that act, safely."; "53% of organizations have already had an agent exceed its permissions." with no source; "Measured results" of −70.6% policy failures and +40.4% compliant completions on τ²-bench and −20.5% agent-model spend per successful task on CarBench, with no link to a method, paper or repo; a Python sample ("from salus import Salus" … "salus.protect(...)" around a do_refund tool) with no install command, package or GitHub link; a 59-second airline-booking demo; no customer logos or testimonials.Source
    • News coverageY Combinator: Salus is in the Winter 2026 batch with a team of 2 and the description "Guardrails to validate your agent's actions before they execute". Its launch post claimed 58% of blocked actions recovered and completed correctly, "up to 60% lower cost" on τ²-bench and a 52% average misalignment reduction on ODCV-Bench across 12 frontier models, with onboarding as a pip install plus a few lines of code.Source
    • News coverageStartupHub review of Salus: integration as "change the URL your agent hits"; it warns that a policy proxy that adds 50ms per action is "dead on arrival", calls the moat "not yet durable", and notes the benchmark method and the 58% figure are not explained.Source
    • Your websiteManifesto: "Agents don't crash. They commit."; "Monitoring is mourning. The dashboard is the obituary. Salus is the intervention."; Salus "returns a reason, not a refusal, because a system that only says no gets removed in a week"; the failure is framed as correctness at execution time, not permissions.Source
  • 04

    The homepage shows "from salus import Salus" and a salus.protect wrapper around a refund tool, but no install command, package link or repository link. The site links no GitHub at all. The launch post sold onboarding as "a pip install plus a few lines of code", yet a developer reading the sample can't tell which package to install.

    Medium impactLow effort First result in 1 week Founder & CEO

    First action

    Add the install command above the homepage sample.

    How to measure it

    Install-line copy events

    Now
    –
    Target
    Tracked from day one; 5% of homepage visitors
    By
    30 days

    Steps

    1. 1

      Confirm the package

      Name and current version of the public SDK.

      Founder & CEO 1 hour

    2. 2

      Add the line

      Install command and copy button above the sample.

      Founder & CEO 2 hours

    3. 3

      Link out

      Package page, quickstart and repo if public.

      Founder & CEO 1 hour

    4. 4

      Explain the key

      One line on the environment variable and how to get it.

      Head of Growth 1 hour

    5. 5

      Repeat on other pages

      Product and integrations.

      Founder & CEO 2 hours

    6. 6

      Track copies

      Copy-button events in analytics.

      Head of Growth 1 hour

    First result
    1 week
    Cost
    $0
    Owner
    Founder & CEO

    Smallest test

    Add the install line to the homepage today.

    Why it works

    Add the exact install command above the sample, plus a copy button, links to the package page and the quickstart, and one line on what the sample needs to run (the environment variable and how to get it). Do the same on the product and integrations pages. If a public SDK repo exists, link it in the header next to Docs.

    Developers judge a tool by whether they can run it in five minutes. A visible install line turns a reading visit into a first try and gives assistants and search engines the package name to match.

    Risks and how to handle them

    • Developers install but can't get a key without applying

      Pair with the shadow-only starter plan (idea 12) or say clearly how keys are issued.

    Evidence

    • Your websiteHomepage (checked November 10, 2026): title "Salus — Runtime authorization for AI agents", H1 "Agents that act, safely."; "53% of organizations have already had an agent exceed its permissions." with no source; "Measured results" of −70.6% policy failures and +40.4% compliant completions on τ²-bench and −20.5% agent-model spend per successful task on CarBench, with no link to a method, paper or repo; a Python sample ("from salus import Salus" … "salus.protect(...)" around a do_refund tool) with no install command, package or GitHub link; a 59-second airline-booking demo; no customer logos or testimonials.Source
    • News coverageY Combinator: Salus is in the Winter 2026 batch with a team of 2 and the description "Guardrails to validate your agent's actions before they execute". Its launch post claimed 58% of blocked actions recovered and completed correctly, "up to 60% lower cost" on τ²-bench and a 52% average misalignment reduction on ODCV-Bench across 12 frontier models, with onboarding as a pip install plus a few lines of code.Source
    • Your websiteTeam page: a Founder & CEO who built the enforcement runtime, policy engine and SDKs (Stanford Mathematics and Computer Science), and a Head of Growth with two decades of cybersecurity go-to-market (ScanAlert through its acquisition by McAfee, then McAfee and TrustedSite); a "Work with us" note but no careers page; the site links only LinkedIn, with no X, GitHub or YouTube.Source
  • 05

    The site title says "Runtime authorization for AI agents" and the H1 says "Agents that act, safely." The YC profile still says "Guardrails to validate your agent's actions". Meanwhile the manifesto argues this is "a correctness problem, not a security or permissions problem", and Permit.io already sells "Runtime Access Control for the AI Era". So a buyer meets three names for one product, and one of them puts Salus in an access-control category its own manifesto rejects.

    Medium impactLow effort First result in 3 weeks Head of Growth

    First action

    Write three candidate category lines and test them in the BDR's first 50 emails.

    How to measure it

    Outbound reply rate for the chosen line

    Now
    –
    Target
    Best line beats current by 20%
    By
    45 days

    Steps

    1. 1

      Draft names

      Three category lines drawn from the manifesto.

      Head of Growth 2 hours

    2. 2

      Test in outbound

      Split the BDR's first sequences.

      Founding GTM / BDR 2 weeks

    3. 3

      Pick one

      By reply and meeting rate.

      Founder & CEO 1 day

    4. 4

      Update the site

      Title, meta description, H1 support line.

      Head of Growth 2 hours

    5. 5

      Update profiles

      YC profile, LinkedIn, launch page where editable.

      Head of Growth 2 hours

    6. 6

      Brief the team

      One sentence everyone uses.

      Head of Growth 1 hour

    First result
    3 weeks
    Cost
    $0
    Owner
    Head of Growth

    Smallest test

    Update the YC profile line to match the site this week.

    Why it works

    Choose the name that carries the manifesto's idea, for example "action correctness" or "pre-execution checks for agent actions". Then use it in the site title and meta description, the H1 support line, the YC profile, LinkedIn and the BDR's sequences. Keep "authorization" as a secondary keyword on one comparison page (idea 11) rather than the headline.

    A new category is won by repeating one name until buyers use it. A name that sets Salus apart from access control also answers "how is this different from Permit?" before anyone asks.

    Risks and how to handle them

    • Losing searches for "authorization"

      Keep the term in body copy and on a comparison page.

    Evidence

    • Your websiteHomepage (checked November 10, 2026): title "Salus — Runtime authorization for AI agents", H1 "Agents that act, safely."; "53% of organizations have already had an agent exceed its permissions." with no source; "Measured results" of −70.6% policy failures and +40.4% compliant completions on τ²-bench and −20.5% agent-model spend per successful task on CarBench, with no link to a method, paper or repo; a Python sample ("from salus import Salus" … "salus.protect(...)" around a do_refund tool) with no install command, package or GitHub link; a 59-second airline-booking demo; no customer logos or testimonials.Source
    • Your websiteManifesto: "Agents don't crash. They commit."; "Monitoring is mourning. The dashboard is the obituary. Salus is the intervention."; Salus "returns a reason, not a refusal, because a system that only says no gets removed in a week"; the failure is framed as correctness at execution time, not permissions.Source
    • News coverageY Combinator: Salus is in the Winter 2026 batch with a team of 2 and the description "Guardrails to validate your agent's actions before they execute". Its launch post claimed 58% of blocked actions recovered and completed correctly, "up to 60% lower cost" on τ²-bench and a 52% average misalignment reduction on ODCV-Bench across 12 frontier models, with onboarding as a pip install plus a few lines of code.Source
    • Competitor researchArcade's homepage: "Govern every action for every AI agent", free to start with "Try Free" self-serve signup; Permit.io's homepage: "Runtime Access Control for the AI Era" with an MCP Gateway and a "Start building" signup; RuntimeWire reports Arcjet launched agent runtime security on September 17, 2026, with deterministic checks in application code that block tool, API and database calls before execution.Source

This quarter 13

The core plays for the next three months.

  • 06

    The integrations page names 15 frameworks, voice platforms and protocols, but none has its own page, and every link goes to the same few docs pages. The 10-URL sitemap has no integration pages. Developers search for their own stack ("LangGraph tool call validation", "Vapi function call guardrails"), and the founding engineer is being hired to deploy Salus into exactly those stacks.

    High impactMedium effort First result in 6-8 weeks Founding SWE / MLE

    First action

    Pick the four stacks that come up most in pilot applications.

    How to measure it

    Applications with an integration page as entry

    Now
    –
    Target
    20% of applications
    By
    1 quarter

    Steps

    1. 1

      Rank stacks

      From pilot form answers and deployments.

      Founder & CEO 1 day

    2. 2

      Write a template

      Code, pattern, held-action example, next step.

      Head of Growth 2 days

    3. 3

      Build four pages

      LangGraph, MCP, Vapi, Retell.

      Founding SWE / MLE 2 weeks

    4. 4

      Test the code

      Every snippet runs as written.

      Founding SWE / MLE 2 days

    5. 5

      Add to sitemap

      And link from /integrations.

      Founder & CEO 1 hour

    6. 6

      Extend

      Two more stacks a month.

      Founding SWE / MLE monthly

    7. 7

      Measure

      Search impressions and applications by page.

      Head of Growth monthly

    First result
    6-8 weeks
    Cost
    Engineering and writing time
    Owner
    Founding SWE / MLE

    Smallest test

    Publish the LangGraph page first and watch its search impressions for a month.

    Why it works

    Build one page per integration, starting with LangGraph, MCP, Vapi and Retell. Each page gets the 10-line integration for that stack, the pattern it uses (SDK, webhook or executor), one held-action example in that stack's terms, and a link to the quickstart. Write each page from a real deployment the founding engineer has done.

    Integration pages match the searches developers make when they are already building, and they show that Salus fits without a rewrite, which is the homepage's main promise.

    Risks and how to handle them

    • Snippets go stale as frameworks change

      Run snippets in CI and date each page.

    Builds on

    Evidence

    • Your websiteIntegrations page: frameworks LangChain, LangGraph, CrewAI, AutoGen, PydanticAI, LlamaIndex and Google ADK; voice and realtime Vapi, Retell, LiveKit Agents and Pipecat; MCP, HTTP, computer use and custom tool loops; three patterns (SDK, webhook for Vapi, Retell and HTTP, executor). None has its own page; all links go to the docs.Source
    • Your websiteSitemap: 10 URLs (home, product, integrations, pricing, manifesto, trust, team, pilot, privacy, terms), all last modified August 21, 2026 except terms (September 3); no blog, case study, use-case, comparison, integration or careers pages. The "Sandbox" menu item redirects to app.usesalus.ai and is not in the sitemap.Source
    • HiringFounding SWE / MLE (YC jobs, $250K-$500K, 0.50%-2.00% equity): "Build the core enforcement engine" that "evaluates policy deterministically"; "Go into customer codebases and get Salus deployed against real agent stacks" such as LangChain, LangGraph and MCP; research on "benchmarks, adversarial evaluation, and the harder problems around agent authorization and provenance"; low-latency, correctness-critical code.Source
  • 07

    The product already imports "recent runs" to map tools and writes, and stress-tests policies "against prior actions and synthetic variants". That means Salus can show a prospect their own failure rate before any integration. The new BDR is paid per qualified meeting that holds and needs a reason for a busy engineering lead to take the call. Today the only demo is a fictional airline booking.

    High impactMedium effort First result in 1 month Founding GTM / BDR

    First action

    Run the replay on one existing pilot's data and turn it into a report template.

    How to measure it

    Replay-to-meeting rate

    Now
    –
    Target
    40% of replays become a held meeting
    By
    1 quarter

    Steps

    1. 1

      Define inputs

      Accepted log formats and the redaction guide.

      Founder & CEO 2 days

    2. 2

      Build the report

      Counts by check family, three receipts, fixes.

      Founding SWE / MLE 1 week

    3. 3

      Write the data terms

      Retention, deletion and use, from the trust page.

      Head of Growth 1 day

    4. 4

      Add to outbound

      The offer as the call to action in sequences.

      Founding GTM / BDR 2 days

    5. 5

      Run replays

      Within 48 hours of receiving data.

      Founding SWE / MLE ongoing

    6. 6

      Walk through

      Founder presents the report on the call.

      Founder & CEO ongoing

    7. 7

      Track

      Offers, replays, meetings, pilots.

      Founding GTM / BDR weekly

    First result
    1 month
    Cost
    Engineering time per replay
    Owner
    Founding GTM / BDR

    Smallest test

    Offer the replay to ten warm contacts this month.

    Why it works

    Package the replay as the outbound offer. The prospect shares an export of a few hundred recent tool calls (redacted, kept under the trust page's retention rules), and Salus returns a two-page report: actions checked, how many would have been held by check family, three sample decision receipts, and the fix the agent would have received. The meeting walks through their report, and the next step is shadow mode on the worst route.

    A finding in the buyer's own data is far more persuasive than a generic demo, and it qualifies the lead: teams with no held actions are not ready, and teams with many are.

    Risks and how to handle them

    • Prospects won't share logs

      Accept a small redacted sample or run it on their side with a local script.

    • Replays take too long

      Cap at 500 actions and standard formats.

    Evidence

    • Your websiteProduct page: 14 check families (from replay and duplicate control to evidence and facts, approvals and receipts); a five-step decision boundary (Proposal, Evaluation, Reasoning, Control, Record); a control plane with Inventory ("Import recent runs and map the tools, writes, and providers in use"), Policies, Stress tests ("against prior actions and synthetic variants"), Decisions in shadow or enforcement, and Recovery; decision receipts with policy version, evidence used, facts missing and timing, exportable as CSV, JSON, agent JSON or Markdown. The only worked use case is airline booking, and Incident Rewind is not explained.Source
    • HiringFounding GTM / BDR, AI Agent Infrastructure (YC jobs, contract, $35-$45 an hour, SF preferred): "Run cold outbound" to a defined best-fit customer profile, use warm intros, "Book qualified meetings" handed to the founder for demo and close, draft LinkedIn and X posts in the founder's voice, own list-building, sequencing and follow-up; a paid one-week trial to book 5 qualified meetings and draft 5 posts; a bonus per qualified meeting that holds; "the product works" and "the benchmarks are strong".Source
    • Your websiteTrust page: "SOC 2 Type II in progress" (no auditor or date); encryption in transit and at rest; isolated workspaces that "fail closed"; configurable retention and redaction; the standard path does not need raw audio, full transcripts or card data; "BAAs for workloads involving PHI"; security, privacy and subprocessor information on request by email, with no subprocessor list or downloadable report.Source
  • 08

    The BDR will own list-building and sequencing for a "defined best-fit customer profile", and the site defines that best-fit customer profile only implicitly: agents whose tool calls change state (a refund in the code sample, a booking in the demo) on the 15 stacks Salus supports. The teams that need Salus most say so in public, when they announce agent features that take actions or post jobs for agent engineers on LangGraph, MCP or Vapi.

    High impactMedium effort First result in 3 weeks Founding GTM / BDR

    First action

    Write the three trigger searches and pull the first week's list.

    How to measure it

    Held meetings from trigger lists

    Now
    –
    Target
    5 a week
    By
    1 quarter

    Steps

    1. 1

      Define triggers

      Action-taking launches, framework job posts, voice-platform builders.

      Head of Growth 1 day

    2. 2

      Pull lists

      Weekly, with the action type noted per account.

      Founding GTM / BDR weekly

    3. 3

      Write openers

      One per action type: refunds, bookings, scheduling, writes.

      Founding GTM / BDR 2 days

    4. 4

      Offer the replay

      As the single call to action.

      Founding GTM / BDR ongoing

    5. 5

      Hand off

      Held meetings to the founder with notes.

      Founding GTM / BDR ongoing

    6. 6

      Review

      Which triggers produce meetings that hold.

      Head of Growth weekly

    First result
    3 weeks
    Cost
    BDR time plus a data tool subscription
    Owner
    Founding GTM / BDR

    Smallest test

    Send 50 trigger-based emails and 50 generic ones and compare held meetings.

    Why it works

    Set up weekly trigger lists from three public sources: product launches of agents that act (refunds, bookings, scheduling, CRM updates), job posts naming supported frameworks plus words like "tool calling" or "production agents", and companies building on the voice platforms Salus integrates with. Each email opens with the specific action type they ship and offers the shadow replay (idea 7). Score lists by meetings that hold, not by reply rate.

    Trigger-based outbound reaches teams at the moment the risk becomes real, and it gives a new BDR a repeatable system instead of a static list.

    Risks and how to handle them

    • Generic outreach to engineers gets ignored

      Name the exact action type and stack in line one.

    Builds on

    Evidence

    • HiringFounding GTM / BDR, AI Agent Infrastructure (YC jobs, contract, $35-$45 an hour, SF preferred): "Run cold outbound" to a defined best-fit customer profile, use warm intros, "Book qualified meetings" handed to the founder for demo and close, draft LinkedIn and X posts in the founder's voice, own list-building, sequencing and follow-up; a paid one-week trial to book 5 qualified meetings and draft 5 posts; a bonus per qualified meeting that holds; "the product works" and "the benchmarks are strong".Source
    • Your websiteIntegrations page: frameworks LangChain, LangGraph, CrewAI, AutoGen, PydanticAI, LlamaIndex and Google ADK; voice and realtime Vapi, Retell, LiveKit Agents and Pipecat; MCP, HTTP, computer use and custom tool loops; three patterns (SDK, webhook for Vapi, Retell and HTTP, executor). None has its own page; all links go to the docs.Source
    • Your websiteHomepage (checked November 10, 2026): title "Salus — Runtime authorization for AI agents", H1 "Agents that act, safely."; "53% of organizations have already had an agent exceed its permissions." with no source; "Measured results" of −70.6% policy failures and +40.4% compliant completions on τ²-bench and −20.5% agent-model spend per successful task on CarBench, with no link to a method, paper or repo; a Python sample ("from salus import Salus" … "salus.protect(...)" around a do_refund tool) with no install command, package or GitHub link; a 59-second airline-booking demo; no customer logos or testimonials.Source
  • 09

    The BDR will draft LinkedIn and X posts in the founder's voice. Salus produces something few vendors have: a receipt for every held action, showing the policy, the evidence used, the facts missing and the fix. The manifesto lines ("Agents don't crash. They commit.") are already strong hooks.

    Medium impactLow effort First result in 1 month Founding GTM / BDR

    First action

    Turn the airline demo's receipt into the first post.

    How to measure it

    Post-sourced pilot applications

    Now
    –
    Target
    4 a month
    By
    1 quarter

    Steps

    1. 1

      Pick receipts

      Benchmark runs or opted-in pilots only.

      Founder & CEO weekly

    2. 2

      Anonymize

      Remove names, amounts and identifiers.

      Founding GTM / BDR weekly

    3. 3

      Design a template

      Receipt image plus three lines.

      Founding GTM / BDR 1 day

    4. 4

      Post

      LinkedIn and X from the founder's account.

      Founding GTM / BDR weekly

    5. 5

      Add a source question

      "Where did you hear about us?" on /pilot.

      Head of Growth 1 hour

    6. 6

      Review

      Which receipts drive applications.

      Head of Growth monthly

    First result
    1 month
    Cost
    BDR time
    Owner
    Founding GTM / BDR

    Smallest test

    Post four receipts in a month and count replies and applications.

    Why it works

    Each week, take one held action (from Salus's own benchmark runs or an opted-in pilot), strip identifying details, and post the receipt as an image with three lines: what the agent tried, what was missing, what it did after the fix. Close with the shadow replay offer. Keep a public archive on the site later.

    Concrete near-misses are shareable and teach the category better than abstract claims. They also give the BDR a steady stream of posts that point to one offer.

    Risks and how to handle them

    • A customer recognizes their data

      Use benchmark runs or written opt-in only.

    Evidence

    • HiringFounding GTM / BDR, AI Agent Infrastructure (YC jobs, contract, $35-$45 an hour, SF preferred): "Run cold outbound" to a defined best-fit customer profile, use warm intros, "Book qualified meetings" handed to the founder for demo and close, draft LinkedIn and X posts in the founder's voice, own list-building, sequencing and follow-up; a paid one-week trial to book 5 qualified meetings and draft 5 posts; a bonus per qualified meeting that holds; "the product works" and "the benchmarks are strong".Source
    • Your websiteProduct page: 14 check families (from replay and duplicate control to evidence and facts, approvals and receipts); a five-step decision boundary (Proposal, Evaluation, Reasoning, Control, Record); a control plane with Inventory ("Import recent runs and map the tools, writes, and providers in use"), Policies, Stress tests ("against prior actions and synthetic variants"), Decisions in shadow or enforcement, and Recovery; decision receipts with policy version, evidence used, facts missing and timing, exportable as CSV, JSON, agent JSON or Markdown. The only worked use case is airline booking, and Incident Rewind is not explained.Source
    • Your websiteManifesto: "Agents don't crash. They commit."; "Monitoring is mourning. The dashboard is the obituary. Salus is the intervention."; Salus "returns a reason, not a refusal, because a system that only says no gets removed in a week"; the failure is framed as correctness at execution time, not permissions.Source
  • 10

    The product page has one worked example, an airline booking, and the homepage sample protects a refund. The sitemap has no use-case pages. Buyers think in their own action ("our agent issues refunds"), and the 14 check families map neatly onto those actions: evidence for amounts, duplicate control for double bookings, approvals for large refunds.

    High impactMedium effort First result in 6-8 weeks Head of Growth

    First action

    Write the refunds page from the homepage do_refund sample.

    How to measure it

    Applications from use-case pages

    Now
    –
    Target
    25% of applications
    By
    1 quarter

    Steps

    1. 1

      Choose four actions

      From pilot answers to "What does your agent do?".

      Head of Growth 1 day

    2. 2

      Map checks

      Failure modes to check families per action.

      Founder & CEO 2 days

    3. 3

      Write pages

      Failure, check, receipt, repair, integration.

      Head of Growth 3 weeks

    4. 4

      Add receipts

      One realistic sample receipt per page.

      Founding SWE / MLE 1 week

    5. 5

      Publish and link

      From the product page and in outbound.

      Head of Growth 1 day

    6. 6

      Measure

      Traffic and applications by page.

      Head of Growth monthly

    First result
    6-8 weeks
    Cost
    Writing time
    Owner
    Head of Growth

    Smallest test

    Publish the refunds page and use it in 30 outbound emails.

    Why it works

    Write four pages, each built the same way: the action, the three ways it goes wrong, the checks that catch each one, a sample receipt and repair, and the integration for the usual stack. Start with refunds (the homepage sample) and bookings (the demo), then voice scheduling and CRM or database writes.

    Use-case pages turn a horizontal tool into a direct answer for each buyer, and the BDR can link the one that matches each prospect's action type.

    Risks and how to handle them

    • Pages read as generic

      Every page shows a real receipt and code.

    Evidence

    • Your websiteProduct page: 14 check families (from replay and duplicate control to evidence and facts, approvals and receipts); a five-step decision boundary (Proposal, Evaluation, Reasoning, Control, Record); a control plane with Inventory ("Import recent runs and map the tools, writes, and providers in use"), Policies, Stress tests ("against prior actions and synthetic variants"), Decisions in shadow or enforcement, and Recovery; decision receipts with policy version, evidence used, facts missing and timing, exportable as CSV, JSON, agent JSON or Markdown. The only worked use case is airline booking, and Incident Rewind is not explained.Source
    • Your websiteHomepage (checked November 10, 2026): title "Salus — Runtime authorization for AI agents", H1 "Agents that act, safely."; "53% of organizations have already had an agent exceed its permissions." with no source; "Measured results" of −70.6% policy failures and +40.4% compliant completions on τ²-bench and −20.5% agent-model spend per successful task on CarBench, with no link to a method, paper or repo; a Python sample ("from salus import Salus" … "salus.protect(...)" around a do_refund tool) with no install command, package or GitHub link; a 59-second airline-booking demo; no customer logos or testimonials.Source
    • Your websiteSitemap: 10 URLs (home, product, integrations, pricing, manifesto, trust, team, pilot, privacy, terms), all last modified August 21, 2026 except terms (September 3); no blog, case study, use-case, comparison, integration or careers pages. The "Sandbox" menu item redirects to app.usesalus.ai and is not in the sitemap.Source
  • 11

    Arcade, Permit.io and Arcjet all now promise control over agent actions, each from a different angle (an actions runtime, access control with an MCP gateway, deterministic in-code checks). A public review notes that teams can also build guardrails themselves. Salus has no comparison page, yet its manifesto already states the difference: permission is not correctness.

    Medium impactMedium effort First result in 6 weeks Head of Growth

    First action

    Draft the approach comparison table from the manifesto.

    How to measure it

    Comparison-page applications

    Now
    –
    Target
    5 a month
    By
    1 quarter

    Steps

    1. 1

      Draft the table

      Approaches, questions answered, placement, receipts, repair.

      Head of Growth 1 day

    2. 2

      Check facts

      Each vendor's current site, dated.

      Head of Growth 1 day

    3. 3

      Write the approach page

      Category level, no attacks.

      Head of Growth 1 week

    4. 4

      Write vendor pages

      Where they differ and where they combine.

      Head of Growth 1 week

    5. 5

      Review claims

      Founder checks technical accuracy.

      Founder & CEO 1 day

    6. 6

      Refresh

      Quarterly.

      Head of Growth quarterly

    First result
    6 weeks
    Cost
    Writing time
    Owner
    Head of Growth

    Smallest test

    Publish the approach page alone and track its traffic.

    Why it works

    Write one page that compares approaches rather than attacking vendors: what each approach answers ("may this agent call this tool?" against "is this specific action grounded in current evidence?"), where each sits, latency, receipts and repair. Then write short, fair "Salus and Permit.io" and "Salus and Arcade" pages, showing where they work together, with dated facts from each vendor's own site.

    Evaluators search for alternatives and comparisons before they shortlist. A fair, approach-level page earns trust and puts Salus's framing at the center.

    Risks and how to handle them

    • Outdated or unfair claims

      Date pages, cite sources, invite corrections.

    Builds on

    Evidence

    • Competitor researchArcade's homepage: "Govern every action for every AI agent", free to start with "Try Free" self-serve signup; Permit.io's homepage: "Runtime Access Control for the AI Era" with an MCP Gateway and a "Start building" signup; RuntimeWire reports Arcjet launched agent runtime security on September 17, 2026, with deterministic checks in application code that block tool, API and database calls before execution.Source
    • Your websiteManifesto: "Agents don't crash. They commit."; "Monitoring is mourning. The dashboard is the obituary. Salus is the intervention."; Salus "returns a reason, not a refusal, because a system that only says no gets removed in a week"; the failure is framed as correctness at execution time, not permissions.Source
    • News coverageStartupHub review of Salus: integration as "change the URL your agent hits"; it warns that a policy proxy that adds 50ms per action is "dead on arrival", calls the moat "not yet durable", and notes the benchmark method and the 58% figure are not explained.Source
    • Your websiteSitemap: 10 URLs (home, product, integrations, pricing, manifesto, trust, team, pilot, privacy, terms), all last modified August 21, 2026 except terms (September 3); no blog, case study, use-case, comparison, integration or careers pages. The "Sandbox" menu item redirects to app.usesalus.ai and is not in the sitemap.Source
  • 12

    Salus has no free tier or trial: the cheapest plan is $500 a month, reached through an application. Arcade is free to start with a self-serve signup, and Permit.io lets developers start building. Salus's own shadow mode blocks nothing, so it carries little risk for either side, and every route already starts there.

    High impactMedium effort First result in 2 months Founder & CEO

    First action

    Decide the check allowance and the upgrade trigger.

    How to measure it

    Starter accounts upgrading to Pilot

    Now
    –
    Target
    10% within 60 days of signup
    By
    2 quarters

    Steps

    1. 1

      Set limits

      One agent, check allowance, shadow only.

      Founder & CEO 1 day

    2. 2

      Build signup

      Instant workspace and key.

      Founding SWE / MLE 2 weeks

    3. 3

      Weekly report

      Email of would-hold actions with sample receipts.

      Founding SWE / MLE 1 week

    4. 4

      Upgrade path

      One click to Pilot to turn on enforcement.

      Founder & CEO 1 week

    5. 5

      Add to pricing

      A fourth column and FAQ entry.

      Head of Growth 1 day

    6. 6

      Follow up

      BDR contacts starters with many held actions.

      Founding GTM / BDR ongoing

    7. 7

      Measure

      Signups, activation, upgrades.

      Head of Growth monthly

    First result
    2 months
    Cost
    Engineering time and hosting for free checks
    Owner
    Founder & CEO

    Smallest test

    Invite 20 developers from outbound and declined pilot applications to a manually provisioned shadow-only workspace.

    Why it works

    Offer a self-serve shadow-only plan: one agent, a small monthly check allowance (for example 10,000), receipts and the would-hold report, no enforcement and community support. Upgrading to Pilot unlocks enforcement, more agents and Incident Rewind. Put the plan on /pricing with its own instant signup, and keep the founder-led pilot for teams ready to enforce.

    A developer who sees their own would-hold receipts has a reason to ask for budget. Shadow-only limits support load and risk, and it puts Salus on equal footing with self-serve rivals.

    Risks and how to handle them

    • Free users never upgrade

      Keep enforcement paid and tie the weekly report to the upgrade.

    • Support load for a two-person team

      Docs and community support only.

    Builds on

    Evidence

    • Your websitePricing: Pilot "$500 / month", "Month-to-month · Cancel anytime", 10 agents and 200K action checks a month; Partner "$2,500 / month", 50 agents, 1.5M checks, priority support; Enterprise "Custom", unlimited agents, deployment review; "Unlimited seats"; shadow or enforcement per route; Incident Rewind included; checks stop at the ceiling with notice ("Enforcement never silently degrades"); no free tier or trial; every plan's "Get started" goes to /pilot.Source
    • Competitor researchArcade's homepage: "Govern every action for every AI agent", free to start with "Try Free" self-serve signup; Permit.io's homepage: "Runtime Access Control for the AI Era" with an MCP Gateway and a "Start building" signup; RuntimeWire reports Arcjet launched agent runtime security on September 17, 2026, with deterministic checks in application code that block tool, API and database calls before execution.Source
    • Your websiteProduct page: 14 check families (from replay and duplicate control to evidence and facts, approvals and receipts); a five-step decision boundary (Proposal, Evaluation, Reasoning, Control, Record); a control plane with Inventory ("Import recent runs and map the tools, writes, and providers in use"), Policies, Stress tests ("against prior actions and synthetic variants"), Decisions in shadow or enforcement, and Recovery; decision receipts with policy version, evidence used, facts missing and timing, exportable as CSV, JSON, agent JSON or Markdown. The only worked use case is airline booking, and Incident Rewind is not explained.Source
  • 13

    The trust page says SOC 2 Type II is "in progress" with no auditor or date, and offers security, privacy and subprocessor information only by email. A product that sits in front of refunds and writes will face a security review in every serious deal, and the Head of Growth has spent two decades selling security to exactly these reviewers.

    Medium impactLow effort First result in 1 month Head of Growth

    First action

    Publish the subprocessor list and the SOC 2 timeline.

    How to measure it

    Security review cycle time

    Now
    –
    Target
    -30%
    By
    2 quarters

    Steps

    1. 1

      List subprocessors

      Name, purpose, region.

      Founder & CEO 2 hours

    2. 2

      State the SOC 2 timeline

      Audit window and expected report.

      Head of Growth 1 hour

    3. 3

      Draw the data flow

      What leaves the customer and what is stored.

      Founder & CEO 1 day

    4. 4

      Fill a questionnaire

      One standard format, downloadable.

      Head of Growth 3 days

    5. 5

      Add an NDA form

      For the full pack.

      Head of Growth 1 hour

    6. 6

      Link it

      From pricing, pilot and outbound.

      Head of Growth 1 hour

    First result
    1 month
    Cost
    $0 to a few hundred dollars for a trust-page tool
    Owner
    Head of Growth

    Smallest test

    Add the subprocessor list this week.

    Why it works

    Add to /trust: a subprocessor list, an architecture and data-flow diagram, the SOC 2 Type II audit window and expected report date, a pre-filled standard security questionnaire, fail-closed and retention defaults, and a request form for the full pack under NDA. Link it from pricing and the pilot page.

    Self-serve answers shorten the review that often stalls early-stage security deals, and a visible SOC 2 timeline answers the first question before anyone asks.

    Risks and how to handle them

    • Publishing details too early

      Keep architecture high level; send the rest under NDA.

    Evidence

    • Your websiteTrust page: "SOC 2 Type II in progress" (no auditor or date); encryption in transit and at rest; isolated workspaces that "fail closed"; configurable retention and redaction; the standard path does not need raw audio, full transcripts or card data; "BAAs for workloads involving PHI"; security, privacy and subprocessor information on request by email, with no subprocessor list or downloadable report.Source
    • Your websiteTeam page: a Founder & CEO who built the enforcement runtime, policy engine and SDKs (Stanford Mathematics and Computer Science), and a Head of Growth with two decades of cybersecurity go-to-market (ScanAlert through its acquisition by McAfee, then McAfee and TrustedSite); a "Work with us" note but no careers page; the site links only LinkedIn, with no X, GitHub or YouTube.Source
    • Your websitePricing: Pilot "$500 / month", "Month-to-month · Cancel anytime", 10 agents and 200K action checks a month; Partner "$2,500 / month", 50 agents, 1.5M checks, priority support; Enterprise "Custom", unlimited agents, deployment review; "Unlimited seats"; shadow or enforcement per route; Incident Rewind included; checks stop at the ceiling with notice ("Enforcement never silently degrades"); no free tier or trial; every plan's "Get started" goes to /pilot.Source
  • 14

    Every protected action now waits on Salus. A public review of the company says a policy proxy that adds 50ms per action is "dead on arrival". Salus combines deterministic checks with optional reasoning gates, and its FAQ promises "Enforcement never silently degrades". Yet the site publishes no latency figures, uptime or status page. The founding engineer is being hired for "low-latency, correctness-critical" code.

    High impactMedium effort First result in 1 month Founding SWE / MLE

    First action

    Measure p50 and p95 added latency for deterministic checks on the τ²-bench runs.

    How to measure it

    Calls where latency blocks the next step

    Now
    –
    Target
    Under 10%
    By
    2 quarters

    Steps

    1. 1

      Instrument

      Added latency per check family.

      Founding SWE / MLE 1 week

    2. 2

      Benchmark

      p50 and p95, deterministic and reasoning.

      Founding SWE / MLE 1 week

    3. 3

      Document failure modes

      Timeouts, outages, fail closed or open per route.

      Founder & CEO 2 days

    4. 4

      Launch status page

      Uptime and incident history.

      Founding SWE / MLE 1 day

    5. 5

      Publish /performance

      Numbers, method and date.

      Head of Growth 2 days

    6. 6

      Refresh

      Monthly numbers.

      Founding SWE / MLE monthly

    First result
    1 month
    Cost
    Engineering time plus a status page tool
    Owner
    Founding SWE / MLE

    Smallest test

    Publish the deterministic p95 on the product page.

    Build the monitor. A live page fed by the public registries and re-checked every day. Usually live within days.

    Book a working session

    Why it works

    Publish a /performance page with p50 and p95 added latency for deterministic checks and for reasoning gates, measured on the benchmark workloads and stated with the region. Explain what happens on timeout or outage (fail closed, configurable per route) and launch a public status page with uptime history. Put the deterministic p95 next to the homepage code sample.

    Latency and availability are the first objections to anything inline. Publishing numbers turns a fear into a spec, and it pushes engineering to keep them good.

    Risks and how to handle them

    • Reasoning gates look slow

      Show which checks are deterministic by default and when reasoning runs.

    Evidence

    • News coverageStartupHub review of Salus: integration as "change the URL your agent hits"; it warns that a policy proxy that adds 50ms per action is "dead on arrival", calls the moat "not yet durable", and notes the benchmark method and the 58% figure are not explained.Source
    • Your websiteProduct page: 14 check families (from replay and duplicate control to evidence and facts, approvals and receipts); a five-step decision boundary (Proposal, Evaluation, Reasoning, Control, Record); a control plane with Inventory ("Import recent runs and map the tools, writes, and providers in use"), Policies, Stress tests ("against prior actions and synthetic variants"), Decisions in shadow or enforcement, and Recovery; decision receipts with policy version, evidence used, facts missing and timing, exportable as CSV, JSON, agent JSON or Markdown. The only worked use case is airline booking, and Incident Rewind is not explained.Source
    • Your websitePricing: Pilot "$500 / month", "Month-to-month · Cancel anytime", 10 agents and 200K action checks a month; Partner "$2,500 / month", 50 agents, 1.5M checks, priority support; Enterprise "Custom", unlimited agents, deployment review; "Unlimited seats"; shadow or enforcement per route; Incident Rewind included; checks stop at the ceiling with notice ("Enforcement never silently degrades"); no free tier or trial; every plan's "Get started" goes to /pilot.Source
    • HiringFounding SWE / MLE (YC jobs, $250K-$500K, 0.50%-2.00% equity): "Build the core enforcement engine" that "evaluates policy deterministically"; "Go into customer codebases and get Salus deployed against real agent stacks" such as LangChain, LangGraph and MCP; research on "benchmarks, adversarial evaluation, and the harder problems around agent authorization and provenance"; low-latency, correctness-critical code.Source
  • 15

    Every route starts in shadow or enforcement, chosen per route, and the pilot's third step is to "Run the route in shadow or enforcement". The value only lands when a route moves to enforcement. Shadow mode can quietly become the permanent state: the agent keeps making mistakes, the receipts pile up unread, and the $500 plan starts to look optional.

    High impactMedium effort First result in 1 month Founder & CEO

    First action

    Write the readiness rule (minimum checks, maximum false-hold rate).

    How to measure it

    Shadow-to-enforcement within 30 days

    Now
    –
    Target
    60% of pilot routes
    By
    1 quarter

    Steps

    1. 1

      Define readiness

      Minimum checks and false-hold threshold.

      Founder & CEO 1 day

    2. 2

      Add labeling

      Mark holds correct or false from the report.

      Founding SWE / MLE 1 week

    3. 3

      Build the report

      Per route, weekly, by email.

      Founding SWE / MLE 1 week

    4. 4

      One-click switch

      Enforcement from the report.

      Founding SWE / MLE 3 days

    5. 5

      Call stuck routes

      Founder reviews routes not ready after three weeks.

      Founder & CEO weekly

    6. 6

      Track

      Time to enforcement and retention.

      Head of Growth monthly

    First result
    1 month
    Cost
    Engineering time
    Owner
    Founder & CEO

    Smallest test

    Send a manual report to the current pilots this week.

    Why it works

    Send each pilot a weekly report per shadow route: actions checked, would-hold count by check family, a sample of holds for the team to mark correct or false, the false-hold rate, and a clear recommendation ("route refunds is ready: 0 false holds in 412 checks"). Include a one-click switch to enforcement and a short call offer for routes that are not ready.

    A specific, evidence-backed recommendation removes the fear that keeps teams in shadow mode, and enforced routes are what make Salus hard to remove.

    Risks and how to handle them

    • Teams enforce too early and get false holds

      Recommend only when the threshold is met; keep per-route rollback.

    Evidence

    • Your websitePricing: Pilot "$500 / month", "Month-to-month · Cancel anytime", 10 agents and 200K action checks a month; Partner "$2,500 / month", 50 agents, 1.5M checks, priority support; Enterprise "Custom", unlimited agents, deployment review; "Unlimited seats"; shadow or enforcement per route; Incident Rewind included; checks stop at the ceiling with notice ("Enforcement never silently degrades"); no free tier or trial; every plan's "Get started" goes to /pilot.Source
    • Your websitePilot page: "Apply for a founder-led pilot." with a reply promised "within 12 hours"; a four-field form (Name, Work email, Organization, "What does your agent do?"); steps "Map the consequential action route", "Connect your existing agent and tools", "Run the route in shadow or enforcement"; no price, plan choice or pilot length on the page; a "Book a demo" link to a 30-minute calendar slot; "No mailing list."Source
    • Your websiteProduct page: 14 check families (from replay and duplicate control to evidence and facts, approvals and receipts); a five-step decision boundary (Proposal, Evaluation, Reasoning, Control, Record); a control plane with Inventory ("Import recent runs and map the tools, writes, and providers in use"), Policies, Stress tests ("against prior actions and synthetic variants"), Decisions in shadow or enforcement, and Recovery; decision receipts with policy version, evidence used, facts missing and timing, exportable as CSV, JSON, agent JSON or Markdown. The only worked use case is airline booking, and Incident Rewind is not explained.Source
  • 16

    Pilot includes 200K checks and Partner 1.5M, so the next tier gives 7.5 times the checks for 5 times the price. Checks stop at the ceiling, and Salus promises to notify customers before they hit it. That notice already goes to the most engaged accounts at the moment they are using Salus most.

    Medium impactLow effort First result in 1-2 months Head of Growth

    First action

    Draft the 70% email with the monthly value summary.

    How to measure it

    Upgrades after a usage notice

    Now
    –
    Target
    30% of accounts that reach 90%
    By
    2 quarters

    Steps

    1. 1

      Set thresholds

      70% and 90% of checks and agents.

      Founder & CEO 1 day

    2. 2

      Summarize value

      Checked, held, repaired, top receipts.

      Founding SWE / MLE 3 days

    3. 3

      Write emails

      Two notices with the upgrade math.

      Head of Growth 1 day

    4. 4

      Add upgrade link

      Prorated Partner or a check pack.

      Founder & CEO 2 days

    5. 5

      Alert the founder

      At 90% for a call.

      Founder & CEO 1 hour

    6. 6

      Measure

      Upgrade rate per notice.

      Head of Growth monthly

    First result
    1-2 months
    Cost
    $0
    Owner
    Head of Growth

    Smallest test

    Send the value summary manually to the first account that passes 70%.

    Why it works

    Send the notice at 70% and 90% of the allowance. Include what Salus did that month (actions checked, held, repaired, and the top three receipts) and a one-click upgrade to Partner, prorated, plus an option to add a one-time check pack. Copy the founder on the 90% notice so they can call.

    A usage warning that shows value converts better than a bare limit, and Partner's better unit price makes the upgrade easy to justify.

    Risks and how to handle them

    • Customers feel pressured

      Lead with value and offer the check pack as a smaller step.

    Evidence

    • Your websitePricing: Pilot "$500 / month", "Month-to-month · Cancel anytime", 10 agents and 200K action checks a month; Partner "$2,500 / month", 50 agents, 1.5M checks, priority support; Enterprise "Custom", unlimited agents, deployment review; "Unlimited seats"; shadow or enforcement per route; Incident Rewind included; checks stop at the ceiling with notice ("Enforcement never silently degrades"); no free tier or trial; every plan's "Get started" goes to /pilot.Source
    • Your websiteProduct page: 14 check families (from replay and duplicate control to evidence and facts, approvals and receipts); a five-step decision boundary (Proposal, Evaluation, Reasoning, Control, Record); a control plane with Inventory ("Import recent runs and map the tools, writes, and providers in use"), Policies, Stress tests ("against prior actions and synthetic variants"), Decisions in shadow or enforcement, and Recovery; decision receipts with policy version, evidence used, facts missing and timing, exportable as CSV, JSON, agent JSON or Markdown. The only worked use case is airline booking, and Incident Rewind is not explained.Source
  • 17

    Salus supports Vapi, Retell, LiveKit Agents and Pipecat, with a webhook pattern for Vapi and Retell. Its trust page says the standard path doesn't need raw audio or full transcripts, which matters for voice teams. Voice agents book appointments, take payments and update records in real time, with no human watching the screen, which is exactly the consequential-action problem.

    High impactMedium effort First result in 2 months Head of Growth

    First action

    Write the Vapi guide with working code and send it to their developer relations team.

    How to measure it

    Voice-platform-sourced applications

    Now
    –
    Target
    10 a quarter
    By
    2 quarters

    Steps

    1. 1

      Pick two platforms

      Where most voice pilots run.

      Founder & CEO 1 day

    2. 2

      Write guides

      Code, receipt, data-minimization note.

      Founding SWE / MLE 2 weeks

    3. 3

      Pitch

      Developer relations at each platform.

      Head of Growth 1 week

    4. 4

      Publish jointly

      On both sites where possible.

      Head of Growth 2 weeks

    5. 5

      Run a session

      Webinar or livestream build.

      Founder & CEO 1 day

    6. 6

      List the integration

      In each platform's directory if one exists.

      Head of Growth 1 week

    7. 7

      Measure

      Tracked links to /pilot.

      Head of Growth monthly

    First result
    2 months
    Cost
    Staff time
    Owner
    Head of Growth

    Smallest test

    Publish one voice guide and post it in the platform's community channel.

    Why it works

    Approach each voice platform's developer relations team with a ready-made, co-branded guide: "Stop wrong bookings and payments in your voice agent", including working webhook code, a sample receipt and the data-minimization note. Offer to run a joint webinar or livestream, and ask to be listed in whatever integration or partner directory each platform keeps.

    Platforms want their builders to ship safely, and a guide that brings its own code costs them nothing. For Salus it is distribution to developers who are building action-taking agents right now.

    Risks and how to handle them

    • Platforms are slow to respond

      Publish the guide on Salus's own site first and share it with their community.

    Builds on

    Evidence

    • Your websiteIntegrations page: frameworks LangChain, LangGraph, CrewAI, AutoGen, PydanticAI, LlamaIndex and Google ADK; voice and realtime Vapi, Retell, LiveKit Agents and Pipecat; MCP, HTTP, computer use and custom tool loops; three patterns (SDK, webhook for Vapi, Retell and HTTP, executor). None has its own page; all links go to the docs.Source
    • Your websiteTrust page: "SOC 2 Type II in progress" (no auditor or date); encryption in transit and at rest; isolated workspaces that "fail closed"; configurable retention and redaction; the standard path does not need raw audio, full transcripts or card data; "BAAs for workloads involving PHI"; security, privacy and subprocessor information on request by email, with no subprocessor list or downloadable report.Source
    • Your websiteDocs (fetched twice on November 10, 2026): docs.usesalus.ai, /docs/cloud-quickstart, /docs/integration and /docs/cloud-custom each return only the page title "Salus Cloud" to a reader that does not run JavaScript; docs.usesalus.ai has no sitemap.xml, robots.txt or llms.txt (all 404). The homepage, product and integrations pages send developers to these docs ("Read the quickstart", "Python integration", "Webhook integration").Source
  • 18

    Salus's public launch so far was the YC launch page. The site links only LinkedIn, with no GitHub, X or YouTube, and the BDR will draft X posts for an account the site doesn't link. Developer tools spread where developers talk, and the manifesto lines are made for that audience.

    Medium impactMedium effort First result in 1 day Founder & CEO

    First action

    Add GitHub and X links to the site footer.

    How to measure it

    Launch-week applications and starter signups

    Now
    –
    Target
    50
    By
    Launch week

    Steps

    1. 1

      Link profiles

      GitHub and X in the footer.

      Head of Growth 1 hour

    2. 2

      Ready the path

      Readable docs, install line, starter or demo.

      Founder & CEO 2 weeks

    3. 3

      Write the post

      Demo, receipt, code, latency.

      Founder & CEO 1 day

    4. 4

      Prepare answers

      Latency, fail-closed, data handling, pricing.

      Head of Growth 1 day

    5. 5

      Launch

      Founder answers comments all day.

      Founder & CEO 1 day

    6. 6

      Follow up

      BDR contacts commenters who build agents.

      Founding GTM / BDR 1 week

    First result
    1 day
    Cost
    $0
    Owner
    Founder & CEO

    Smallest test

    Share the demo in two agent-builder communities first and refine the pitch.

    Why it works

    Once ideas 1 and 4 are live (and ideally the starter plan, idea 12), post a Show HN built around the playable demo: a stale-fare booking held and repaired, the receipt, the 10-line integration, and the published latency numbers. Link GitHub and X from the site footer before launch, and have the founder answer every comment the same day.

    A technical audience rewards a concrete demo and open numbers, and the discussion itself becomes a public FAQ that search and assistants index.

    Risks and how to handle them

    • Harsh comments on benchmarks

      Publish the method first (idea 3).

    Builds on

    Evidence

    • Your websiteTeam page: a Founder & CEO who built the enforcement runtime, policy engine and SDKs (Stanford Mathematics and Computer Science), and a Head of Growth with two decades of cybersecurity go-to-market (ScanAlert through its acquisition by McAfee, then McAfee and TrustedSite); a "Work with us" note but no careers page; the site links only LinkedIn, with no X, GitHub or YouTube.Source
    • HiringFounding GTM / BDR, AI Agent Infrastructure (YC jobs, contract, $35-$45 an hour, SF preferred): "Run cold outbound" to a defined best-fit customer profile, use warm intros, "Book qualified meetings" handed to the founder for demo and close, draft LinkedIn and X posts in the founder's voice, own list-building, sequencing and follow-up; a paid one-week trial to book 5 qualified meetings and draft 5 posts; a bonus per qualified meeting that holds; "the product works" and "the benchmarks are strong".Source
    • News coverageY Combinator: Salus is in the Winter 2026 batch with a team of 2 and the description "Guardrails to validate your agent's actions before they execute". Its launch post claimed 58% of blocked actions recovered and completed correctly, "up to 60% lower cost" on τ²-bench and a 52% average misalignment reduction on ODCV-Bench across 12 frontier models, with onboarding as a pip install plus a few lines of code.Source
    • Your websiteManifesto: "Agents don't crash. They commit."; "Monitoring is mourning. The dashboard is the obituary. Salus is the intervention."; Salus "returns a reason, not a refusal, because a system that only says no gets removed in a week"; the failure is framed as correctness at execution time, not permissions.Source

Big bets 3

Larger plays that can change the trajectory.

  • 19

    Arcjet now offers deterministic checks inside application code, and a public review calls the core proxy "replicable by a competent team in a few months" with a moat "not yet durable". Salus's founding engineer will build an engine that "evaluates policy deterministically". The deterministic layer is the easiest part to copy, while receipts, stress tests against prior runs, reasoning gates and recovery are much harder.

    High impactHigh effort First result in 2 quarters Founder & CEO

    First action

    Decide the open/closed line and the license.

    How to measure it

    Local-to-Cloud conversion

    Now
    –
    Target
    5% of active installs
    By
    3 quarters

    Steps

    1. 1

      Draw the line

      What is open, what stays in Cloud.

      Founder & CEO 1 week

    2. 2

      Pick a license

      Permissive or source-available.

      Founder & CEO 1 week

    3. 3

      Extract the engine

      In-process library and policy format.

      Founding SWE / MLE 6 weeks

    4. 4

      Write examples

      LangGraph, MCP and voice webhook.

      Founding SWE / MLE 2 weeks

    5. 5

      Add the upgrade

      One line to send receipts to Cloud.

      Founding SWE / MLE 2 weeks

    6. 6

      Launch

      Repo, docs, Show HN.

      Founder & CEO 1 week

    7. 7

      Measure

      Stars, installs, Cloud connections.

      Head of Growth monthly

    First result
    2 quarters
    Cost
    One quarter of engineering focus
    Owner
    Founder & CEO

    Smallest test

    Publish the policy format and a validator as open source first.

    Why it works

    Release the deterministic check engine and policy format as an open-source library that runs in-process with no network hop. Salus Cloud then sells what teams can't easily build: the evidence cache across sessions, reasoning gates, receipts with retention and export, stress tests, recovery and the control plane. Make the upgrade one config line.

    Open source wins developer adoption and trust in a crowded category, it answers latency and lock-in objections, and every local install becomes a cloud lead with its usage already visible.

    Risks and how to handle them

    • Competitors fork the engine

      Keep the evidence cache, receipts and reasoning in Cloud; ship faster.

    • Maintenance load for a small team

      Small surface area and clear contribution rules.

    Builds on

    Evidence

    • HiringFounding SWE / MLE (YC jobs, $250K-$500K, 0.50%-2.00% equity): "Build the core enforcement engine" that "evaluates policy deterministically"; "Go into customer codebases and get Salus deployed against real agent stacks" such as LangChain, LangGraph and MCP; research on "benchmarks, adversarial evaluation, and the harder problems around agent authorization and provenance"; low-latency, correctness-critical code.Source
    • Competitor researchArcade's homepage: "Govern every action for every AI agent", free to start with "Try Free" self-serve signup; Permit.io's homepage: "Runtime Access Control for the AI Era" with an MCP Gateway and a "Start building" signup; RuntimeWire reports Arcjet launched agent runtime security on September 17, 2026, with deterministic checks in application code that block tool, API and database calls before execution.Source
    • News coverageStartupHub review of Salus: integration as "change the URL your agent hits"; it warns that a policy proxy that adds 50ms per action is "dead on arrival", calls the moat "not yet durable", and notes the benchmark method and the 58% figure are not explained.Source
  • 20

    Salus already reports results on τ²-bench, CarBench and ODCV-Bench, and the founding engineer role includes research on "benchmarks, adversarial evaluation" and "agent authorization and provenance". But the numbers live as three unexplained figures on a homepage. Nobody yet owns a benchmark for the specific failure Salus sells against: an authorized, well-formed action that is wrong (stale price, wrong recipient, duplicate refund).

    High impactHigh effort First result in 2 quarters Founding SWE / MLE

    First action

    Write the task taxonomy from the 14 check families.

    How to measure it

    External leaderboard submissions

    Now
    –
    Target
    10
    By
    3 quarters

    Steps

    1. 1

      Define tasks

      Action types and traps, from the check families.

      Founding SWE / MLE 3 weeks

    2. 2

      Build the harness

      Reproducible runs and scoring.

      Founding SWE / MLE 6 weeks

    3. 3

      Run baselines

      With and without guard layers.

      Founding SWE / MLE 2 weeks

    4. 4

      Write the paper

      Method, limits, results.

      Founder & CEO 3 weeks

    5. 5

      Publish

      Repo, leaderboard page, paper.

      Founder & CEO 1 week

    6. 6

      Invite submissions

      Framework and vendor teams.

      Head of Growth ongoing

    7. 7

      Refresh

      Quarterly results.

      Founding SWE / MLE quarterly

    First result
    2 quarters
    Cost
    A quarter of research time plus compute
    Owner
    Founding SWE / MLE

    Smallest test

    Publish 50 tasks and one baseline as a preview.

    Why it works

    Build an open benchmark of realistic consequential-action tasks (bookings, refunds, payments, record updates) with traps for stale evidence, missing confirmation and duplicates. Publish the dataset, harness and a leaderboard for agents with and without any guard layer, competitors included. Release a short paper and refresh it each quarter.

    Whoever defines the measurement shapes how buyers frame the problem. A neutral, open benchmark earns citations, talks and backlinks that marketing alone can't buy.

    Risks and how to handle them

    • Seen as rigged in Salus's favor

      Open data and harness, external reviewers, report Salus's own failures.

    Builds on

    Evidence

    • HiringFounding SWE / MLE (YC jobs, $250K-$500K, 0.50%-2.00% equity): "Build the core enforcement engine" that "evaluates policy deterministically"; "Go into customer codebases and get Salus deployed against real agent stacks" such as LangChain, LangGraph and MCP; research on "benchmarks, adversarial evaluation, and the harder problems around agent authorization and provenance"; low-latency, correctness-critical code.Source
    • News coverageY Combinator: Salus is in the Winter 2026 batch with a team of 2 and the description "Guardrails to validate your agent's actions before they execute". Its launch post claimed 58% of blocked actions recovered and completed correctly, "up to 60% lower cost" on τ²-bench and a 52% average misalignment reduction on ODCV-Bench across 12 frontier models, with onboarding as a pip install plus a few lines of code.Source
    • Your websiteHomepage (checked November 10, 2026): title "Salus — Runtime authorization for AI agents", H1 "Agents that act, safely."; "53% of organizations have already had an agent exceed its permissions." with no source; "Measured results" of −70.6% policy failures and +40.4% compliant completions on τ²-bench and −20.5% agent-model spend per successful task on CarBench, with no link to a method, paper or repo; a Python sample ("from salus import Salus" … "salus.protect(...)" around a do_refund tool) with no install command, package or GitHub link; a 59-second airline-booking demo; no customer logos or testimonials.Source
  • 21

    Every consequential action gets a receipt with the policy version, matched rule, evidence used, facts missing and timing, and receipts export as CSV, JSON or Markdown with configurable retention. Salus already offers BAAs for regulated workloads. Compliance and audit teams in regulated companies need exactly this record before they let agents act, and they buy differently from engineers: on evidence, retention and sign-off.

    High impactHigh effort First result in 2-3 quarters Head of Growth

    First action

    Interview three pilot teams' compliance or risk owners about what an auditor would ask for.

    How to measure it

    Audit add-on contracts

    Now
    –
    Target
    3
    By
    3 quarters

    Steps

    1. 1

      Interview buyers

      Compliance and risk owners at pilots.

      Head of Growth 3 weeks

    2. 2

      Map controls

      Receipt fields to common audit controls.

      Head of Growth 2 weeks

    3. 3

      Build storage

      Long retention and tamper evidence.

      Founding SWE / MLE 6 weeks

    4. 4

      Approval workflow

      Sign-off on policy changes.

      Founding SWE / MLE 4 weeks

    5. 5

      Price it

      Enterprise add-on.

      Founder & CEO 1 week

    6. 6

      Sell

      Two-threaded with engineering and compliance.

      Head of Growth ongoing

    First result
    2-3 quarters
    Cost
    Two quarters of product work
    Owner
    Head of Growth

    Smallest test

    Produce one auditor-style receipt export for a pilot and ask their compliance lead for feedback.

    Why it works

    Create an Enterprise add-on for regulated teams: long retention, tamper-evident receipt storage, auditor-ready exports mapped to common control frameworks, approval workflows for policy changes, and a quarterly attestation report. Sell it to compliance and risk owners alongside the engineering champion, starting with teams already in pilot.

    Audit evidence turns Salus from an engineering tool into a control that compliance requires, which raises the contract value and makes Salus much harder to remove.

    Risks and how to handle them

    • Long regulated sales cycles

      Start with existing pilots and keep the add-on optional.

    • SOC 2 Type II not yet complete

      Time the launch with the report.

    Builds on

    Evidence

    • Your websiteTrust page: "SOC 2 Type II in progress" (no auditor or date); encryption in transit and at rest; isolated workspaces that "fail closed"; configurable retention and redaction; the standard path does not need raw audio, full transcripts or card data; "BAAs for workloads involving PHI"; security, privacy and subprocessor information on request by email, with no subprocessor list or downloadable report.Source
    • Your websiteProduct page: 14 check families (from replay and duplicate control to evidence and facts, approvals and receipts); a five-step decision boundary (Proposal, Evaluation, Reasoning, Control, Record); a control plane with Inventory ("Import recent runs and map the tools, writes, and providers in use"), Policies, Stress tests ("against prior actions and synthetic variants"), Decisions in shadow or enforcement, and Recovery; decision receipts with policy version, evidence used, facts missing and timing, exportable as CSV, JSON, agent JSON or Markdown. The only worked use case is airline booking, and Incident Rewind is not explained.Source
    • Your websitePricing: Pilot "$500 / month", "Month-to-month · Cancel anytime", 10 agents and 200K action checks a month; Partner "$2,500 / month", 50 agents, 1.5M checks, priority support; Enterprise "Custom", unlimited agents, deployment review; "Unlimited seats"; shadow or enforcement per route; Incident Rewind included; checks stop at the ceiling with notice ("Enforcement never silently degrades"); no free tier or trial; every plan's "Get started" goes to /pilot.Source

Wildcards 3

Unexpected experiments, each with a cheap way to test it.

  • 22

    The homepage has no customer logos or testimonials, and the sitemap has no case studies. Every proof point is a benchmark. The BDR is paid for meetings that hold, and warm intros from a team already running Salus hold far better than cold outbound.

    Medium impactLow effort First result in 2 months Founder & CEO

    First action

    Offer the deal to the current pilots this week.

    How to measure it

    Stories plus referred meetings

    Now
    –
    Target
    3 stories and 5 referred meetings
    By
    2 quarters

    Steps

    1. 1

      Write terms

      Credit for a story or an intro.

      Founder & CEO 1 day

    2. 2

      Offer to pilots

      Teams with an enforced route first.

      Founder & CEO 1 week

    3. 3

      Interview

      Problem, route, holds, repairs, latency.

      Head of Growth per story

    4. 4

      Approve

      Customer signs off on wording and naming.

      Head of Growth per story

    5. 5

      Publish /customers

      And add to the sitemap.

      Head of Growth 1 day

    6. 6

      Use in outbound

      Story matched to prospect action type.

      Founding GTM / BDR ongoing

    First result
    2 months
    Cost
    One month of credit per partner
    Owner
    Founder & CEO

    Smallest test

    Ask one pilot for an anonymized story this month.

    Why it works

    Offer the first ten teams that move a route to enforcement a design-partner deal: one month of Pilot credit for an anonymized or named case study (problem, route, holds and repairs, latency), or for an intro to another team shipping action-taking agents. Publish stories on a new /customers page and use them in outbound.

    Early buyers of infrastructure look for peers who went first. A small credit is cheap compared with the trust one real story earns.

    Risks and how to handle them

    • Customers won't be named

      Accept anonymized stories with real numbers.

    Builds on

    Evidence

    • Your websiteHomepage (checked November 10, 2026): title "Salus — Runtime authorization for AI agents", H1 "Agents that act, safely."; "53% of organizations have already had an agent exceed its permissions." with no source; "Measured results" of −70.6% policy failures and +40.4% compliant completions on τ²-bench and −20.5% agent-model spend per successful task on CarBench, with no link to a method, paper or repo; a Python sample ("from salus import Salus" … "salus.protect(...)" around a do_refund tool) with no install command, package or GitHub link; a 59-second airline-booking demo; no customer logos or testimonials.Source
    • Your websiteSitemap: 10 URLs (home, product, integrations, pricing, manifesto, trust, team, pilot, privacy, terms), all last modified August 21, 2026 except terms (September 3); no blog, case study, use-case, comparison, integration or careers pages. The "Sandbox" menu item redirects to app.usesalus.ai and is not in the sitemap.Source
    • HiringFounding GTM / BDR, AI Agent Infrastructure (YC jobs, contract, $35-$45 an hour, SF preferred): "Run cold outbound" to a defined best-fit customer profile, use warm intros, "Book qualified meetings" handed to the founder for demo and close, draft LinkedIn and X posts in the founder's voice, own list-building, sequencing and follow-up; a paid one-week trial to book 5 qualified meetings and draft 5 posts; a bonus per qualified meeting that holds; "the product works" and "the benchmarks are strong".Source
  • 23

    The manifesto's central line, "Agents don't crash. They commit.", describes a kind of failure that keeps showing up in public: agents that refund, book or send things they shouldn't. Salus has 14 named check families but no blog or content section to connect them to real events. The sitemap has 10 pages and no content at all.

    Medium impactMedium effort First result in 2 months Head of Growth

    First action

    Collect ten documented public cases and map each to a check family.

    How to measure it

    Referring domains to the library

    Now
    –
    Target
    25
    By
    2 quarters

    Steps

    1. 1

      Collect cases

      Reputable public sources only.

      Head of Growth 2 weeks

    2. 2

      Write a format

      Summary, action, missing evidence, check, repair.

      Head of Growth 1 day

    3. 3

      Map checks

      Founder reviews each mapping.

      Founder & CEO weekly

    4. 4

      Publish

      /incidents with tags by action type.

      Head of Growth 1 week

    5. 5

      Open submissions

      A simple form with source required.

      Head of Growth 1 day

    6. 6

      Promote

      One entry per week in founder posts.

      Founding GTM / BDR weekly

    First result
    2 months
    Cost
    Writing time
    Owner
    Head of Growth

    Smallest test

    Publish the first ten entries on one page.

    Why it works

    Start a public library of documented agent action failures from reputable public reports. For each entry, give a neutral summary, the action type, the missing evidence or rule, the Salus check family that applies and what the agent would have been told. Add one or two entries a week, let readers submit cases, and tag entries by action type so the use-case pages (idea 10) can link to them.

    A neutral, well-sourced library becomes a reference that journalists, researchers and buyers cite, and every entry is a concrete argument for checks at execution time.

    Risks and how to handle them

    • Naming companies invites disputes

      Neutral summaries of public reports only, with links and no speculation.

    Evidence

    • Your websiteManifesto: "Agents don't crash. They commit."; "Monitoring is mourning. The dashboard is the obituary. Salus is the intervention."; Salus "returns a reason, not a refusal, because a system that only says no gets removed in a week"; the failure is framed as correctness at execution time, not permissions.Source
    • Your websiteProduct page: 14 check families (from replay and duplicate control to evidence and facts, approvals and receipts); a five-step decision boundary (Proposal, Evaluation, Reasoning, Control, Record); a control plane with Inventory ("Import recent runs and map the tools, writes, and providers in use"), Policies, Stress tests ("against prior actions and synthetic variants"), Decisions in shadow or enforcement, and Recovery; decision receipts with policy version, evidence used, facts missing and timing, exportable as CSV, JSON, agent JSON or Markdown. The only worked use case is airline booking, and Incident Rewind is not explained.Source
    • Your websiteSitemap: 10 URLs (home, product, integrations, pricing, manifesto, trust, team, pilot, privacy, terms), all last modified August 21, 2026 except terms (September 3); no blog, case study, use-case, comparison, integration or careers pages. The "Sandbox" menu item redirects to app.usesalus.ai and is not in the sitemap.Source
  • 24

    The homepage is fast for real visitors: PageSpeed scores it 100 on mobile, and real users see the main content in 749 ms at the 75th percentile with no layout shift. But the page still downloads about 14.6 MB, while its scripts, styles, images and fonts add up to under 0.5 MB, so nearly all of the weight is media such as the demo video. Developers often open links on a phone or a tethered laptop.

    Low impactLow effort First result in 1 day Founder & CEO

    First action

    Set the video to preload=none with a poster image.

    How to measure it

    Homepage download size

    Now
    –
    Target
    Under 1 MB
    By
    1 week

    Steps

    1. 1

      Confirm the source

      Check which file makes up the 14.6 MB.

      Founder & CEO 30 minutes

    2. 2

      Add a poster

      Still frame with a play button.

      Founder & CEO 1 hour

    3. 3

      Load on click

      preload=none or a click-to-load player.

      Founder & CEO 1 hour

    4. 4

      Add captions

      For muted viewing.

      Head of Growth 1 hour

    5. 5

      Track plays

      Play events as an intent signal.

      Head of Growth 1 hour

    6. 6

      Re-test

      PageSpeed and transfer size.

      Founder & CEO 15 minutes

    First result
    1 day
    Cost
    $0
    Owner
    Founder & CEO

    Smallest test

    Set preload=none today and re-run PageSpeed.

    Why it works

    Show a poster image with a play button and load the video only on click (or use a streaming host). Keep the 59-second length and add captions. Re-test to confirm the page drops below 1 MB.

    It keeps today's excellent speed scores and makes the first visit cheap on mobile data. It also makes "play" a measurable intent signal for the demo.

    Risks and how to handle them

    • Fewer people watch

      Compare play rate before and after; the poster can show the held $299 fare to invite a click.

    Evidence

    • Google PageSpeedHomepage test (November 10, 2026): mobile performance 100, and accessibility, best practices and SEO 100; real visitors pass with LCP 749 ms at the 75th percentile and no layout shift. The page still downloads about 14.6 MB in 19 requests, while scripts, styles, images and fonts add up to under 0.5 MB, so most of the weight is media such as the demo video; Google Tag Manager is the largest script (178 KB).Source
    • Your websiteHomepage (checked November 10, 2026): title "Salus — Runtime authorization for AI agents", H1 "Agents that act, safely."; "53% of organizations have already had an agent exceed its permissions." with no source; "Measured results" of −70.6% policy failures and +40.4% compliant completions on τ²-bench and −20.5% agent-model spend per successful task on CarBench, with no link to a method, paper or repo; a Python sample ("from salus import Salus" … "salus.protect(...)" around a do_refund tool) with no install command, package or GitHub link; a 59-second airline-booking demo; no customer logos or testimonials.Source

03

How Programmable can help

Where an idea in this playbook is something Programmable builds or runs, it's listed here with what we'd do. Everything else is yours to run.

Have it built

ProgrammableAgency

1 of the 24 ideas are things Programmable builds and runs for its customers: pages, outreach, prospect lists, booking flows and monitors. Pick one, or send the whole list.

Monitors1A live page fed by the public registries and re-checked every day.

A 30-minute session with the engineer who would build it, not a sales call. You leave with a plan whether or not you hire us.

Book a working session

Usually live within days.

Tools for the rest

ProgrammableDigital

Show it, don't just describe it

These ideas need something to show: a demo, a video, an animation, graphics or a deck. Digital turns screenshots or the site into all of them, in the company's own brand.

Make a demo (opens programmable.digital in a new tab)
More example playbooks and reviews

More growth playbooks

Other companies we have researched, each read from what they are hiring for.

  • MedeAnalytics growth strategyEnterprise healthcare analytics and performance improvement: the Health Fabric data platform, payer and provider analytics, the Avi AI…21 growth ideas
  • Subject growth strategyVideo-first digital curriculum for grades 6 to 12, sold to schools and districts with state-certified teachers of record, credit recovery…21 growth ideas
  • Fuel GTM growth strategyA founder-led GTM engineering and RevOps consultancy that builds and connects CRMs (Salesforce, HubSpot), Clay enrichment and research…23 growth ideas
  • Omniscient Digital growth strategyAn organic growth agency for B2B software companies: SEO, answer engine optimization (AEO), content, digital PR, link building, CRO and…21 growth ideas
  • SplitMetrics growth strategyMobile marketing software and services: SplitMetrics Acquire for Apple Ads management with the Iris AI agent, App Radar for ASO, Optimize…21 growth ideas
  • Iron Will Outfitters growth strategyEngineers and sells premium fixed-blade broadheads (Single Bevel, Wide, S, V, SLR, crossbow, turkey and small game), arrow components…20 growth ideas
  • Storetasker growth strategyA marketplace of vetted freelance Shopify developers, designers and marketers for ecommerce brands, for projects and monthly retainers…21 growth ideas
  • Spring Fertility growth strategyA physician-led fertility group offering IVF, IUI, egg and embryo freezing, genetic testing, donor and surrogacy services and LGBTQ+…22 growth ideas

See every playbook and website review

Independent research. Programmable wrote this playbook from public sources, as they stood on November 10, 2026, and cites them throughout. Salus did not commission, review or endorse it, and Programmable is not affiliated with Salus. Company and product names are trademarks of their owners, used only to identify them. The ideas are suggestions and opinion, not professional, legal or financial advice, and not statements about how Salus runs its business.

Something wrong, or do you represent Salus? Email [email protected] and we will correct it or take it down.

30/60/90 roadmap

What to focus on in each month, and the ideas that carry it.

  1. Days 1 to 306 ideas

    Readable docs, a priced pilot path, sourced numbers, the install line, one category name and a lighter homepage.

  2. Days 31 to 6015 ideas

    Give the new BDR an offer and a list (shadow replay, trigger lists, receipt posts); build integration, use-case and comparison pages; a starter plan, trust pack, latency page, graduation reports, cap alerts, voice partners, a Show HN, design partners and the incident library.

  3. Days 61 to 903 ideas

    Commit to the big bets: an open local check runtime, a public agent-action benchmark and an audit-evidence package for regulated teams.

The research behind the plan

3 research areas, each summarised in a few lines. Open any of them for the full detail.

Customers

What customers say

Customers value keeping their model, agent loop, framework and backend unchanged. The most cited problem: agents that take authorized, well-formed but wrong actions.

  • Running a route in shadow mode before turning enforcement on
  • Finding out about a bad refund or booking hours later
  • In their words: “consequential action”

Market

Competitive landscape

4 competitors compared: Arcade, Permit.io, Arcjet and In-house guardrails.

  • Gap to exploit: Correctness of an authorized action (evidence, freshness, totals), not just permission
  • Arcade: Broader platform, less focus on evidence-grounded correctness
  • Permit.io: Answers "may this agent do this?", not "is this action correct?"

Hiring signals

Salus is a two-person team hiring its first two people, and the roles show the plan: get the product into customers' hands.

  • Founding GTM / BDR - AI Agent Infrastructure: Outbound and founder-voice posts, paid per meeting that holds. Ideas 7, 8, 9.
  • Founding SWE / MLE: Deterministic engine, customer deployments, benchmark research. Ideas 6, 14, 19, 20.

From the team that wrote this playbook

All products
  • ProgrammableAgencyWorking AI systems in days, not quarters (opens programmable.agency in a new tab)
  • ProgrammableInfoAn agent that finds, researches and asks firstNew accounts get 100,000 free enrichment credits with code BETA100K (opens programmable.info in a new tab)
  • ProgrammableDigitalScreenshots in, an animated product demo out (opens programmable.digital in a new tab)
  • ProgrammableWatchKnow the day a supplier falls out of compliance (opens programmable.watch in a new tab)
  • ProgrammableTalkGet registrants for your events, and know who’s coming (opens programmable.talk in a new tab)
  • ProgrammableMediaAn AI note taker, interpreter and webinar host for your calls (opens programmable.media in a new tab)
  • ProgrammablePageText what you need, your customer gets a page that works (opens programmable.page in a new tab)
  • ProgrammableCommunityForums with labelled AI members, never empty (opens programmable.community in a new tab)
ProgrammableMarketing
ExamplesLeaderboardsProductsTermsPrivacyCookiesDPALegalContact

© 2026 AI-Powered Growth Intelligence