Epsilab growth playbook

Keep the toy as the hero, and turn each nav label into a real, server-rendered page. /environments: each environment family, the software or workflow it simulates, horizon length, reset and verifier type, export formats. /approach: environments from real-life data, stateful resets, graders that catch reward hacking, trajectory QA. /agentic-evolution: the founders' essay. In the header, add a visible "Talk to us" button ([email protected] or a three-question scoping form) and a "Docs · pip install epsilab" link, and link the YC F26 badge to the YC profile. Add every page to the sitemap with Organization and SoftwareApplication markup, and repeat the meta-description claims in the visible text.

Frontier-lab buyers check a vendor's site before a call, and competitors make that easy: HUD's homepage offers "Start building for free", "Book a call" and "pip install hud", with live usage counters. Pages with real text are what search engines and AI assistants can cite and what directories copy into their listings. The work is a few days for a team that already ships a v0.17 SDK.

[ ] /environments with each family, horizon, verifier and export formats
[ ] /approach: real-life data, resets, graders, reward-hacking checks
[ ] /agentic-evolution founders' essay
[ ] Header: Talk to us (form or founders@) and Docs · pip install epsilab
[ ] Sign in links to the app
[ ] YC F26 badge links to the YC profile
[ ] Text in server-rendered HTML
[ ] Sitemap lists every page
[ ] Organization and SoftwareApplication markup
[ ] Meta-description claims repeated in visible text

• Losing the distinctive look

  Keep the toy as the hero and style the new pages to match.

• Revealing lab partners' work

  Describe environment families and methods, never named customers or their tasks.

• Your websiteHomepage (text read on November 9, 2026): H1 "The training grounds for real-world agents.", subhead "RL environments for real-world use cases.", "Backed by Y Combinator", nav labels "Approach", "The Agentic Evolution", "RL Environments" and "Blog", a "Sign in" label, a "press a button" sticker toy and a ticker announcing "A note from the founders". The text read finds no link to docs, the SDK, pricing, contact or a demo, no customer proof and no careers page, although a "join us" sticker is shown.Source
• Your websiteSitemap (checked November 9, 2026): 2 URLs, the homepage and one blog post, "Hello from the Training Grounds" (October 16, 2026), which promises deep dives on building environments and verifiers, notes on long-horizon agents and product updates, and asks readers to suggest topics at [email protected].Source
• Your websiteSearch snippet: title "Epsilab | Long-Horizon RL Environments for Agent Training"; meta description "Run isolated sessions, compare agent performance, and export governed training data in GRPO, DPO, and KTO formats." GRPO, DPO and KTO do not appear in the page's visible text.Source
• PyPIepsilab 0.17.22 (July 27, 2026; first release September 10, 2025; Beta; Apache 2.0): "Python SDK and CLI for the Epsilab RL Environment Hub and Marketplace". It browses, runs and deploys hosted environments, runs interactive sessions and long-horizon episodes, batch-evaluates a policy, exports GRPO, DPO, SFT and KTO data, and publishes environments and tools; it has optional extras for four training stacks. The Evaluations feature is marked deprecated, to be removed on 2026-12-31. It links docs.epsilab.com and a GitHub source repo.Source
• Competitor researchHUD, "The platform for building RL environments": "Start building for free" with no credit card, "Book a call", "pip install hud", GitHub, Discord and docs links, live counters (13,764 environments built, 4,831,034 task runs), a DataVendor marketplace where AI labs buy data, and QA agents that audit traces for grader mistakes and reward hacking.See the research Source

Submit Epsilab to rl-list.com with the Long Horizon, Enterprise and Computer Use tags and a link to the new /environments page. Then email the authors of the 2026 RL environment vendor roundups with a two-line description and the YC F26 profile. Keep one standard description everywhere.

Directories are where a lab's research ops team starts a vendor search, and a listing also gives AI assistants a third-party source to cite.

• Listed with thin information

  Ship the /environments page first so the listing links somewhere useful.

• #1 Give the homepage readable pages behind its nav (RL Environments, Approach, Talk to us) so labs, search engines and AI assistants can see what Epsilab sells

• Competitor researchrl-list.com tracks 42 RL environment vendors (tags: Code 23, Computer Use 17, Enterprise 20, Long Horizon 14); Epsilab is not listed. Long-horizon entries include Bespoke Labs, Proximal, Idler and Prime Intellect.See the research Source
• Y CombinatorEpsilab, YC Fall 2026, active, founded 2026, 4 people, San Francisco; tagline "Long horizon RL environments from real-life data"; described as an applied AI research lab focused on RL for frontier model training, covering real-world environment simulation, long-horizon workflows and multi-player agents. No launch posts. YC lists the X account as EpsilabRL, while other profiles point to an "epsilabai" handle.Source

Write a 600-word dated note: why long-horizon agents fail, why environments built from real-life data, what Epsilab ships today (hub, SDK, graders, export formats) and how to start a pilot. Publish it at /blog, link the ticker to it, and post a shorter version as a YC Launch with a clear ask (labs and agent teams who want a scoping call).

A launch post creates one well-timed burst of attention from an audience that already trusts YC companies, and it gives the ticker somewhere to go.

• Too abstract

  Include one concrete environment and one grader failure it caught.

• Your websiteHomepage (text read on November 9, 2026): H1 "The training grounds for real-world agents.", subhead "RL environments for real-world use cases.", "Backed by Y Combinator", nav labels "Approach", "The Agentic Evolution", "RL Environments" and "Blog", a "Sign in" label, a "press a button" sticker toy and a ticker announcing "A note from the founders". The text read finds no link to docs, the SDK, pricing, contact or a demo, no customer proof and no careers page, although a "join us" sticker is shown.Source
• Y CombinatorEpsilab, YC Fall 2026, active, founded 2026, 4 people, San Francisco; tagline "Long horizon RL environments from real-life data"; described as an applied AI research lab focused on RL for frontier model training, covering real-world environment simulation, long-horizon workflows and multi-player agents. No launch posts. YC lists the X account as EpsilabRL, while other profiles point to an "epsilabai" handle.Source
• Your websiteSitemap (checked November 9, 2026): 2 URLs, the homepage and one blog post, "Hello from the Training Grounds" (October 16, 2026), which promises deep dives on building environments and verifiers, notes on long-horizon agents and product updates, and asks readers to suggest topics at [email protected].Source

Update the Product Hunt product page to the RL environments positioning (or start a new product and mark the old one discontinued), link it to epsilab.com, and ask the directory sites still showing the trading description to update or remove it. Use one X handle everywhere: YC lists EpsilabRL, while other profiles point to an epsilabai handle. Add a single line to /approach on how the quant work led to environments built from real-life data, so the history reads as a strength.

Vendor checks are quick, and conflicting descriptions slow down trust. Turned into a story, the pivot shows the team has shipped products before.

• Directories don't respond

  Outrank them with the new pages and listings.

• Old listingsProduct Hunt still lists Epsilab as "Cursor for Quants" (turn trading ideas into algorithms in plain language), linking to epsilab.ai, with 7 upvotes and no reviews; a chatgate.ai profile updated September 5, 2026 still calls it a platform that turns plain-English trading ideas into quant strategies. epsilab.ai now 308-redirects to epsilab.com, and the homepage headline in January 2026 was "Quant Trading Platform".Source
• Y CombinatorEpsilab, YC Fall 2026, active, founded 2026, 4 people, San Francisco; tagline "Long horizon RL environments from real-life data"; described as an applied AI research lab focused on RL for frontier model training, covering real-world environment simulation, long-horizon workflows and multi-player agents. No launch posts. YC lists the X account as EpsilabRL, while other profiles point to an "epsilabai" handle.Source

Either make the Apache-2.0 SDK repo public (it is already open-licensed on PyPI) with a README, quickstart and two example environments, or change the project URLs in the next release to the docs. Pin a README on the GitHub organization that says what Epsilab is and links the site, docs and PyPI.

Developers judge an SDK by its repo. Public code, issues and examples cost little and build trust with the engineers who recommend vendors inside labs.

• Exposing internal code

  Publish a clean copy of the client SDK only, after a review.

• PyPIepsilab 0.17.22 (July 27, 2026; first release September 10, 2025; Beta; Apache 2.0): "Python SDK and CLI for the Epsilab RL Environment Hub and Marketplace". It browses, runs and deploys hosted environments, runs interactive sessions and long-horizon episodes, batch-evaluates a policy, exports GRPO, DPO, SFT and KTO data, and publishes environments and tools; it has optional extras for four training stacks. The Evaluations feature is marked deprecated, to be removed on 2026-12-31. It links docs.epsilab.com and a GitHub source repo.Source
• GitHubThe source repo linked from PyPI (EpsilabAI/epsilab-python) returns 404, and the EpsilabAI organization "has no public repositories"; its listed contact is [email removed].Source
• Competitor researchPrime Intellect's Environments Hub: 2,500+ community environments built on its open-source verifiers library, with "Browse the hub", GitHub and docs links.See the research Source

Write a one-page guide mapping each Evaluations call to its replacement (batch evaluation across provisioned sessions), with a before-and-after code sample. Show a runtime warning that links to it, email every account that called Evaluations in the last 90 days, and offer a 20-minute migration call.

A clean deprecation keeps early users, and with only a few lab accounts each one matters. The call is also a chance to learn what they use next.

• A user can't migrate in time

  Offer a short extension for named accounts.

• PyPIepsilab 0.17.22 (July 27, 2026; first release September 10, 2025; Beta; Apache 2.0): "Python SDK and CLI for the Epsilab RL Environment Hub and Marketplace". It browses, runs and deploys hosted environments, runs interactive sessions and long-horizon episodes, batch-evaluates a policy, exports GRPO, DPO, SFT and KTO data, and publishes environments and tools; it has optional extras for four training stacks. The Evaluations feature is marked deprecated, to be removed on 2026-12-31. It links docs.epsilab.com and a GitHub source repo.Source

Split the sticker toy's code so it loads after the headline, preload the pixel font (or render the headline as HTML text), and remove the render-blocking requests. Profile the button and sticker handlers that cause slow interactions. Add a favicon and fix the one ARIA accessibility issue.

Lab researchers often open vendor sites on their phones from a chat link, and a faster, cleaner first view supports the precision the brand promises.

• The toy feels slower to appear

  Show a static sticker image until the code loads.

• PageSpeedMobile 75/100, desktop 89. Lab mobile LCP 3.9s; the LCP element is pixel-font text inside an SVG; 693 KB of JavaScript in a 1,079 KB page, with "reduce unused JavaScript" worth about 2.55s (519 KB) and render-blocking requests about 0.45s. Real Chrome visitors (75th percentile): LCP 1.5s and CLS 0, both good, but INP 328 ms, above the 200 ms good line. No favicon; one accessibility issue (prohibited ARIA attributes).Source

Publish a catalogue page per environment family: the software or workflow it simulates, typical horizon length (steps and hours), state and reset model, grader type, a public sample task with a recorded baseline trajectory, and the SDK command to run it. Gate full task sets behind a scoping call.

Researchers can judge realism and difficulty in minutes, which shortens the evaluation step of a lab deal, and the sample tasks double as SDK onboarding.

• Sample tasks leak into training sets and lose value

  Publish samples only; keep held-out task sets private.

• #1 Give the homepage readable pages behind its nav (RL Environments, Approach, Talk to us) so labs, search engines and AI assistants can see what Epsilab sells

• PyPIepsilab 0.17.22 (July 27, 2026; first release September 10, 2025; Beta; Apache 2.0): "Python SDK and CLI for the Epsilab RL Environment Hub and Marketplace". It browses, runs and deploys hosted environments, runs interactive sessions and long-horizon episodes, batch-evaluates a policy, exports GRPO, DPO, SFT and KTO data, and publishes environments and tools; it has optional extras for four training stacks. The Evaluations feature is marked deprecated, to be removed on 2026-12-31. It links docs.epsilab.com and a GitHub source repo.Source
• Competitor researchHUD, "The platform for building RL environments": "Start building for free" with no credit card, "Book a call", "pip install hud", GitHub, Discord and docs links, live counters (13,764 environments built, 4,831,034 task runs), a DataVendor marketplace where AI labs buy data, and QA agents that audit traces for grader mistakes and reward hacking.See the research Source
• Competitor researchPrime Intellect's Environments Hub: 2,500+ community environments built on its open-source verifiers library, with "Browse the hub", GitHub and docs links.See the research Source

Define one pilot package: a single environment built from a real workflow the lab picks, 50 to 100 tasks with verifiers, a grader audit for shortcuts and reward hacking, a baseline run with two of the lab's models, and an export in their training format. Fixed price, fixed two-week timeline, credited against a larger contract. Publish the outline on the site.

Fixed scope and price shorten procurement, and a delivered environment with clear results is the best argument for the follow-on contract.

• Scope creep

  Write change requests as paid add-ons.

• HiringTechnical Strategic Projects Lead (San Francisco or remote US, $105K-$145K, 0.50%-1.00% equity): Epsilab builds environments that simulate enterprise software, tools and professional workflows and calls itself "a fast-growing startup working with frontier AI labs". The role owns projects from scoping to delivery and quality checks, builds automation for environment creation, task generation and evaluation, works with frontier lab partners, coordinates engineers and domain experts, and reviews agent trajectories; finance, consulting and investing experience are not required.Source
• HiringMember of Technical Staff, Research Engineer (RL) (San Francisco, $125K-$300K, 0.75%-2.00% equity): builds stateful environments with resets and verifiable outcomes, reward functions and graders that catch shortcuts and reward hacking, experiments on task difficulty, generalization and long-horizon reliability, trajectory analysis that separates model failures from task, tooling or grading problems, and task pipelines with programmatic and expert quality checks.Source

Package three offers: hub access (usage-based sessions and trajectory export), environment licences (a fixed annual fee per environment family with held-out task sets), and custom builds (the pilot, then a statement of work). Publish the structure, and 'from' prices where comfortable, on a /pricing page, with Talk to us for custom work.

Clear packages let self-serve buyers start small and give labs a frame for a bigger contract, so the team stops reinventing each proposal.

• Public prices anchor lab deals low

  Publish structure and entry prices only; keep lab pricing custom.

• #9 A fixed-scope two-week pilot: one custom environment, a grader audit and a trajectory export

• PyPIepsilab 0.17.22 (July 27, 2026; first release September 10, 2025; Beta; Apache 2.0): "Python SDK and CLI for the Epsilab RL Environment Hub and Marketplace". It browses, runs and deploys hosted environments, runs interactive sessions and long-horizon episodes, batch-evaluates a policy, exports GRPO, DPO, SFT and KTO data, and publishes environments and tools; it has optional extras for four training stacks. The Evaluations feature is marked deprecated, to be removed on 2026-12-31. It links docs.epsilab.com and a GitHub source repo.Source
• Your websiteHomepage (text read on November 9, 2026): H1 "The training grounds for real-world agents.", subhead "RL environments for real-world use cases.", "Backed by Y Combinator", nav labels "Approach", "The Agentic Evolution", "RL Environments" and "Blog", a "Sign in" label, a "press a button" sticker toy and a ticker announcing "A note from the founders". The text read finds no link to docs, the SDK, pricing, contact or a demo, no customer proof and no careers page, although a "join us" sticker is shown.Source
• Competitor researchHUD, "The platform for building RL environments": "Start building for free" with no credit card, "Book a call", "pip install hud", GitHub, Discord and docs links, live counters (13,764 environments built, 4,831,034 task runs), a DataVendor marketplace where AI labs buy data, and QA agents that audit traces for grader mistakes and reward hacking.See the research Source

Open a free tier through the SDK: sign up, run the public sample tasks from the catalogue, and export a capped number of trajectories in any supported format. Show a 'first run in five minutes' quickstart on the site. Accounts that hit the cap get an email from the Strategic Projects Lead offering the pilot.

Engineers at labs and agent startups prefer to try before they take a call, and usage reveals which teams are serious.

• Compute abuse

  Tight caps and email verification.

• #8 A public environment catalogue with one runnable sample task per environment

• Competitor researchHUD, "The platform for building RL environments": "Start building for free" with no credit card, "Book a call", "pip install hud", GitHub, Discord and docs links, live counters (13,764 environments built, 4,831,034 task runs), a DataVendor marketplace where AI labs buy data, and QA agents that audit traces for grader mistakes and reward hacking.See the research Source
• PyPIepsilab 0.17.22 (July 27, 2026; first release September 10, 2025; Beta; Apache 2.0): "Python SDK and CLI for the Epsilab RL Environment Hub and Marketplace". It browses, runs and deploys hosted environments, runs interactive sessions and long-horizon episodes, batch-evaluates a policy, exports GRPO, DPO, SFT and KTO data, and publishes environments and tools; it has optional extras for four training stacks. The Evaluations feature is marked deprecated, to be removed on 2026-12-31. It links docs.epsilab.com and a GitHub source repo.Source
• Your websiteHomepage (text read on November 9, 2026): H1 "The training grounds for real-world agents.", subhead "RL environments for real-world use cases.", "Backed by Y Combinator", nav labels "Approach", "The Agentic Evolution", "RL Environments" and "Blog", a "Sign in" label, a "press a button" sticker toy and a ticker announcing "A note from the founders". The text read finds no link to docs, the SDK, pricing, contact or a demo, no customer proof and no careers page, although a "join us" sticker is shown.Source

Once a month, publish a teardown of one anonymized reward hack or grader failure from Epsilab's own environments: what the agent did, why the grader missed it, the fix, and a reusable check. Cross-post short versions to X and LinkedIn and collect them into a public checklist.

Researchers trust vendors who show their failure analysis, and the posts show grader quality, the hardest thing to prove in a sales call. They also feed the blog the founders already promised.

• Exposing partner work

  Recreate every case in a public sample environment.

• HiringMember of Technical Staff, Research Engineer (RL) (San Francisco, $125K-$300K, 0.75%-2.00% equity): builds stateful environments with resets and verifiable outcomes, reward functions and graders that catch shortcuts and reward hacking, experiments on task difficulty, generalization and long-horizon reliability, trajectory analysis that separates model failures from task, tooling or grading problems, and task pipelines with programmatic and expert quality checks.Source
• Your websiteSitemap (checked November 9, 2026): 2 URLs, the homepage and one blog post, "Hello from the Training Grounds" (October 16, 2026), which promises deep dives on building environments and verifiers, notes on long-horizon agents and product updates, and asks readers to suggest topics at [email protected].Source
• Y CombinatorYC's reinforcement learning directory lists 49 companies, including recent environment builders such as Reific (time-series RL environments on live data), Ooak Data (company data into agent test environments), Petrarch (court records into RL environments), Markov (expert computer-use datasets) and Tokenless (detecting reward hacking).See the research Source

Choose one public environment, run the leading open and commercial agents on its held-out tasks, and publish a leaderboard with success rate by horizon length, cost per success and failure categories. Accept submissions through the SDK and refresh it monthly.

Leaderboards earn links, citations and repeat visits, and every submission is an SDK install from exactly the right audience.

• Labs dislike public rankings

  Use public agents only and never results from partner work.

• #8 A public environment catalogue with one runnable sample task per environment

• Your websiteSearch snippet: title "Epsilab | Long-Horizon RL Environments for Agent Training"; meta description "Run isolated sessions, compare agent performance, and export governed training data in GRPO, DPO, and KTO formats." GRPO, DPO and KTO do not appear in the page's visible text.Source
• HiringMember of Technical Staff, Research Engineer (RL) (San Francisco, $125K-$300K, 0.75%-2.00% equity): builds stateful environments with resets and verifiable outcomes, reward functions and graders that catch shortcuts and reward hacking, experiments on task difficulty, generalization and long-horizon reliability, trajectory analysis that separates model failures from task, tooling or grading problems, and task pipelines with programmatic and expert quality checks.Source

Write one page and one runnable notebook per supported training stack: install the extra, run an environment, export trajectories, start a training run. Submit each to the framework's community examples or integrations list, and add an 'Works with' row on the homepage.

Integration pages catch high-intent searches and give each framework's community a reason to link to Epsilab. The engineering work is already done.

• Framework changes break notebooks

  Pin versions and test monthly.

• #5 Fix the PyPI page's dead source link: publish the SDK repo or point it to the docs

• PyPIepsilab 0.17.22 (July 27, 2026; first release September 10, 2025; Beta; Apache 2.0): "Python SDK and CLI for the Epsilab RL Environment Hub and Marketplace". It browses, runs and deploys hosted environments, runs interactive sessions and long-horizon episodes, batch-evaluates a policy, exports GRPO, DPO, SFT and KTO data, and publishes environments and tools; it has optional extras for four training stacks. The Evaluations feature is marked deprecated, to be removed on 2026-12-31. It links docs.epsilab.com and a GitHub source repo.Source

Package two or three finance environments from the team's own stack: a research-to-backtest workflow, a portfolio rebalancing task with verifiable P&L and risk limits, and a long-horizon analyst workflow. Graders score outcome and process. Lead the /environments page and outreach to labs' finance-agent teams with them.

It is a differentiated, defensible line built on code and know-how competitors don't have, and verifiable financial outcomes make strong rewards.

• Market data licence limits

  Use licensable or synthetic data for training environments.

• Crowded finance niche

  Focus on trading and portfolio workflows rather than diligence.

• #8 A public environment catalogue with one runnable sample task per environment

• Old listingsProduct Hunt still lists Epsilab as "Cursor for Quants" (turn trading ideas into algorithms in plain language), linking to epsilab.ai, with 7 upvotes and no reviews; a chatgate.ai profile updated September 5, 2026 still calls it a platform that turns plain-English trading ideas into quant strategies. epsilab.ai now 308-redirects to epsilab.com, and the homepage headline in January 2026 was "Quant Trading Platform".Source
• HiringTechnical Strategic Projects Lead (San Francisco or remote US, $105K-$145K, 0.50%-1.00% equity): Epsilab builds environments that simulate enterprise software, tools and professional workflows and calls itself "a fast-growing startup working with frontier AI labs". The role owns projects from scoping to delivery and quality checks, builds automation for environment creation, task generation and evaluation, works with frontier lab partners, coordinates engineers and domain experts, and reviews agent trajectories; finance, consulting and investing experience are not required.Source
• Competitor researchHalluminate, a data research lab for knowledge-work environments starting with financial services and consulting: its Westworld Finance Diligence Bench has 88 problems from anonymized real deals written by practicing deal professionals; it runs an Expert Network and Data Partnerships and has announced a Series A.See the research Source

Attach a standard report to every delivered environment and each monthly refresh: task difficulty spread, pass rates for baseline agents, share of failures by cause (model, task, tooling, grader), reward hacks found and fixed, and tasks retired. Review it with the lab in a monthly 30-minute call that also scopes the next batch.

It makes quality visible, gives the lab's researcher something to forward internally, and turns delivery into a recurring relationship.

• Showing flaws scares buyers

  Pair every flaw with the fix, which builds trust.

• #9 A fixed-scope two-week pilot: one custom environment, a grader audit and a trajectory export

• HiringMember of Technical Staff, Research Engineer (RL) (San Francisco, $125K-$300K, 0.75%-2.00% equity): builds stateful environments with resets and verifiable outcomes, reward functions and graders that catch shortcuts and reward hacking, experiments on task difficulty, generalization and long-horizon reliability, trajectory analysis that separates model failures from task, tooling or grading problems, and task pipelines with programmatic and expert quality checks.Source
• HiringTechnical Strategic Projects Lead (San Francisco or remote US, $105K-$145K, 0.50%-1.00% equity): Epsilab builds environments that simulate enterprise software, tools and professional workflows and calls itself "a fast-growing startup working with frontier AI labs". The role owns projects from scoping to delivery and quality checks, builds automation for environment creation, task generation and evaluation, works with frontier lab partners, coordinates engineers and domain experts, and reviews agent trajectories; finance, consulting and investing experience are not required.Source

Pay a referral bonus to any domain expert whose referred peer completes a first batch of accepted tasks. Give experts a simple page describing the work, pay and domains wanted, and a personal referral link. Track acceptance rate by referral source.

Practitioners know other practitioners at the right level, and referred experts are usually vetted and faster to onboard.

• Low-quality referrals

  Pay only after accepted work.

• HiringTechnical Strategic Projects Lead (San Francisco or remote US, $105K-$145K, 0.50%-1.00% equity): Epsilab builds environments that simulate enterprise software, tools and professional workflows and calls itself "a fast-growing startup working with frontier AI labs". The role owns projects from scoping to delivery and quality checks, builds automation for environment creation, task generation and evaluation, works with frontier lab partners, coordinates engineers and domain experts, and reviews agent trajectories; finance, consulting and investing experience are not required.Source
• Competitor researchHalluminate, a data research lab for knowledge-work environments starting with financial services and consulting: its Westworld Finance Diligence Bench has 88 problems from anonymized real deals written by practicing deal professionals; it runs an Expert Network and Data Partnerships and has announced a Series A.See the research Source

Invite vetted builders (domain experts, small teams, researchers) to publish long-horizon environments that pass Epsilab's grader and reset checks. Epsilab hosts, QA-certifies and sells access or licences, and pays builders a revenue share. Start with ten invited builders in domains Epsilab doesn't cover.

It widens the catalogue without hiring and makes Epsilab's QA standard the product. That positions it on quality against the larger open hubs.

• Uneven quality hurts the brand

  Certify before listing; delist on failed audits.

• Competes with lab custom work

  Keep held-out sets and custom builds separate.

• #8 A public environment catalogue with one runnable sample task per environment
• #16 Ship every environment with a QA report that splits failures into model, task and grader

• PyPIepsilab 0.17.22 (July 27, 2026; first release September 10, 2025; Beta; Apache 2.0): "Python SDK and CLI for the Epsilab RL Environment Hub and Marketplace". It browses, runs and deploys hosted environments, runs interactive sessions and long-horizon episodes, batch-evaluates a policy, exports GRPO, DPO, SFT and KTO data, and publishes environments and tools; it has optional extras for four training stacks. The Evaluations feature is marked deprecated, to be removed on 2026-12-31. It links docs.epsilab.com and a GitHub source repo.Source
• Competitor researchHUD, "The platform for building RL environments": "Start building for free" with no credit card, "Book a call", "pip install hud", GitHub, Discord and docs links, live counters (13,764 environments built, 4,831,034 task runs), a DataVendor marketplace where AI labs buy data, and QA agents that audit traces for grader mistakes and reward hacking.See the research Source
• Competitor researchPrime Intellect's Environments Hub: 2,500+ community environments built on its open-source verifiers library, with "Browse the hub", GitHub and docs links.See the research Source

Offer operating companies (software vendors, service firms, back offices) a partnership: they share anonymized workflow logs and tool states; Epsilab turns them into environments and pays a licence fee or revenue share, plus a private evaluation of how current agents perform on their work. Start with two partners in one vertical.

Partners get money and an early view of agent readiness for their own workflows, and Epsilab gets exclusive realistic data competitors can't copy.

• Privacy and confidentiality

  Strict anonymization, partner sign-off on every release.

• Y CombinatorEpsilab, YC Fall 2026, active, founded 2026, 4 people, San Francisco; tagline "Long horizon RL environments from real-life data"; described as an applied AI research lab focused on RL for frontier model training, covering real-world environment simulation, long-horizon workflows and multi-player agents. No launch posts. YC lists the X account as EpsilabRL, while other profiles point to an "epsilabai" handle.Source
• Y CombinatorYC's reinforcement learning directory lists 49 companies, including recent environment builders such as Reific (time-series RL environments on live data), Ooak Data (company data into agent test environments), Petrarch (court records into RL environments), Markov (expert computer-use datasets) and Tokenless (detecting reward hacking).See the research Source
• Competitor researchHalluminate, a data research lab for knowledge-work environments starting with financial services and consulting: its Westworld Finance Diligence Bench has 88 problems from anonymized real deals written by practicing deal professionals; it runs an Expert Network and Data Partnerships and has announced a Series A.See the research Source

Once a year, publish a benchmark across several Epsilab environments that measures success against horizon length, recovery from mid-task errors, and cost per completed task, for leading public agents. Release a paper, a public subset and an SDK command to reproduce it. Keep a held-out set private to prevent contamination.

A respected benchmark earns citations, conference talks and inbound from labs that want their models tested, all pointing back to the environments.

• Benchmark contamination

  Private held-out set and rotation each year.

• #13 A public long-horizon leaderboard on one Epsilab environment

• HiringMember of Technical Staff, Research Engineer (RL) (San Francisco, $125K-$300K, 0.75%-2.00% equity): builds stateful environments with resets and verifiable outcomes, reward functions and graders that catch shortcuts and reward hacking, experiments on task difficulty, generalization and long-horizon reliability, trajectory analysis that separates model failures from task, tooling or grading problems, and task pipelines with programmatic and expert quality checks.Source
• Competitor researchHalluminate, a data research lab for knowledge-work environments starting with financial services and consulting: its Westworld Finance Diligence Bench has 88 problems from anonymized real deals written by practicing deal professionals; it runs an Expert Network and Data Partnerships and has announced a Series A.See the research Source

Publish /careers with both roles, the full role text, the hard problems the team works on (long-horizon resets, grader design, failure attribution), how the team works, and links to the teardowns and catalogue. Link it from the join us sticker and point YC posts back to it.

One page helps both hiring and sales: candidates see the problems, and buyers see the depth of the engineering.

• Revealing roadmap

  Describe problems, not partner projects.

• Your websiteHomepage (text read on November 9, 2026): H1 "The training grounds for real-world agents.", subhead "RL environments for real-world use cases.", "Backed by Y Combinator", nav labels "Approach", "The Agentic Evolution", "RL Environments" and "Blog", a "Sign in" label, a "press a button" sticker toy and a ticker announcing "A note from the founders". The text read finds no link to docs, the SDK, pricing, contact or a demo, no customer proof and no careers page, although a "join us" sticker is shown.Source
• HiringMember of Technical Staff, Research Engineer (RL) (San Francisco, $125K-$300K, 0.75%-2.00% equity): builds stateful environments with resets and verifiable outcomes, reward functions and graders that catch shortcuts and reward hacking, experiments on task difficulty, generalization and long-horizon reliability, trajectory analysis that separates model failures from task, tooling or grading problems, and task pipelines with programmatic and expert quality checks.Source
• HiringTechnical Strategic Projects Lead (San Francisco or remote US, $105K-$145K, 0.50%-1.00% equity): Epsilab builds environments that simulate enterprise software, tools and professional workflows and calls itself "a fast-growing startup working with frontier AI labs". The role owns projects from scoping to delivery and quality checks, builds automation for environment creation, task generation and evaluation, works with frontier lab partners, coordinates engineers and domain experts, and reviews agent trajectories; finance, consulting and investing experience are not required.Source

Wire the button to a short, sped-up replay of a real (public sample) trajectory: the environment state, the agent's actions, a mid-task mistake and recovery, and the grader's pass or fail with the reason. End with 'Run this yourself: pip install epsilab' and Talk to us. Keep the sticker art style.

It keeps the brand's playfulness and turns the hero into a ten-second product demo that people share.

• Heavier page

  Load replay data only on press.

• #7 Cut the unused JavaScript and the pixel-font render delay, and add the missing favicon

• Your websiteHomepage (text read on November 9, 2026): H1 "The training grounds for real-world agents.", subhead "RL environments for real-world use cases.", "Backed by Y Combinator", nav labels "Approach", "The Agentic Evolution", "RL Environments" and "Blog", a "Sign in" label, a "press a button" sticker toy and a ticker announcing "A note from the founders". The text read finds no link to docs, the SDK, pricing, contact or a demo, no customer proof and no careers page, although a "join us" sticker is shown.Source
• Your websiteSearch snippet: title "Epsilab | Long-Horizon RL Environments for Agent Training"; meta description "Run isolated sessions, compare agent performance, and export governed training data in GRPO, DPO, and KTO formats." GRPO, DPO and KTO do not appear in the page's visible text.Source
• PageSpeedMobile 75/100, desktop 89. Lab mobile LCP 3.9s; the LCP element is pixel-font text inside an SVG; 693 KB of JavaScript in a 1,079 KB page, with "reduce unused JavaScript" worth about 2.55s (519 KB) and render-blocking requests about 0.45s. Real Chrome visitors (75th percentile): LCP 1.5s and CLS 0, both good, but INP 328 ms, above the 200 ms good line. No favicon; one accessibility issue (prohibited ARIA attributes).Source