Company profile · 3 min read
Cartpole: reinforcement-learning environments with limited public product detail
Cartpole is building reinforcement-learning environments for frontier models, while its current public catalog, pricing and environment examples remain unspecified.
Cofounder of Lazyweb (a16z Speedrun). Previously grew subscriptions and built native ads at Duolingo. Writes First1000 (90,000+ subscribers).
Published · Updated
Cartpole’s current public identity is reinforcement-learning environments for frontier models, but the public evidence is thin. The official profile is current and specific about the category; the homepage is a one-page contact surface, while the most detailed public launch describes an earlier bug-finding product called Jazzberry.
What it does
The YC profile says Cartpole is building reinforcement-learning environments for training frontier models. Its founder, Mateo Perez, is described as a CU Boulder PhD and the company is categorized under AI, reinforcement learning, data labeling and ML. The current homepage only says “cartpole RL environments” and invites contact.
The linked Jazzberry launch is useful history but not evidence that Cartpole currently sells that product. Jazzberry was described as a GitHub pull-request agent that cloned code into a microVM, executed tests and returned a markdown bug report. I am keeping that identity separate from the current Cartpole profile.
Why I’d look closer
The buyer fit is a frontier-model lab or enterprise AI team that needs environments with realistic tools, long-horizon tasks and measurable rewards. The reason to watch Cartpole is the category itself: a model can look strong on static data while failing when it has to explore, recover and complete a multi-step task.
The tradeoff is evidence and packaging. Public pages do not show a catalog, environment examples, pricing, supported simulators or current customer results. That makes a diligence conversation more useful than a confident product comparison. A buyer should also separate a safe sandbox for training from any environment that can execute customer code or reach real systems.
What I’d ask
Which environments exist today, and which are custom builds? What tasks, tools and reward signals do they model? Can a customer inspect the environment code, reset state and reproduce a trajectory? What isolation, data-handling and stop controls apply if the environment executes code or interacts with external systems?
My editorial take
Keep Cartpole on the shortlist for teams building long-horizon model evaluations, but ask for a concrete environment and a customer-owned success metric before treating it as a vendor decision. The company’s current category is clear; the current product proof is not yet public enough to recommend beyond a scoped conversation.
Quick facts
| Field | Sourced detail |
|---|---|
| Current product direction | Reinforcement-learning environments for frontier models |
| Buyer | Frontier AI labs and enterprise teams training/evaluating agents |
| Public availability | Homepage is a contact surface; catalog and pricing not published |
| Historical launch | Jazzberry bug-finding agent; not treated as current Cartpole product evidence |
| Main question | What environment can a buyer inspect, run and measure today? |
Sources checked
| Source | Checked |
|---|---|
| YC company profile | 2026-09-19 |
| Cartpole homepage | 2026-09-19 |
| Jazzberry launch linked from YC | 2026-09-19 |
Cohort context
Cartpole is listed in Spring 2025. In our 2026-09-18 directory snapshot, 97 of 143 listed companies in that cohort have YC’s primary industry label B2B (67.8%). This is a current-directory comparison, not an original intake count or a performance ranking. Nine-cohort dataset.
Public website snapshot
Observed 2026-09-19T16:15:33.304Z in raw homepage HTML. This records visible metadata and advertised links, not agent execution or product quality.
| Signal | Homepage observation |
|---|---|
| Product description metadata | Observed |
| Canonical link | Not observed in this response |
| H1 or H2 heading | Observed |
| Typed structured data | Not observed in this response |
| Docs/developer link | Not observed in this response |
| Pricing link | Not observed in this response |
| llms.txt link | Not observed in this response |
| Markdown alternate | Not observed in this response |
Public observations · Collection method. Missing links here do not establish that a capability or file is absent elsewhere.
About the author
Cofounder of Lazyweb (a16z Speedrun). Previously grew subscriptions and built native ads at Duolingo. Writes First1000 (90,000+ subscribers).
I cofound Lazyweb and publish Mudpie. This is an owner-written publication, not an independent testing organization. Research notes distinguish observations, sourced reporting and editorial judgment.
First1000 ↗ · X ↗