Company profile · 3 min read
HABIT: autonomous research agents for world-model teams
HABIT is building autonomous research agents for teams developing frontier world models, with an inference platform linked from its public site.
Cofounder of Lazyweb (a16z Speedrun). Previously grew subscriptions and built native ads at Duolingo. Writes First1000 (90,000+ subscribers).
Published · Updated
HABIT is building autonomous research agents for teams training frontier world models. The buyer is a model lab, robotics company or applied-AI team that needs agents to generate data, run experiments and improve a world model without wiring every research loop by hand.
What it does
The YC profile describes HABIT as accelerating world models with autonomous research agents. The current homepage uses the same positioning and links to an inference platform, but the checked public pages do not publish a detailed product workflow, pricing, benchmark result or customer case study.
That makes the product direction clearer than the current buying path. A world-model team might use agents to design environments, run simulations, evaluate behavior or search for better representations, but those are reasonable diligence questions rather than published HABIT capabilities. The public profile identifies Michael Trehan as a former software engineer at Radiant Nuclear and investment banker at JP Morgan, and Harish Palani as a former AI/ML research scientist at Amazon and deep-learning lead at Twitch.
Why I’d look closer
The opportunity is in the research loop around models, not another model endpoint. If an agent can choose experiments, execute them and retain the useful result, a small team can explore more of a world-model design space. The tradeoff is evidence: world-model work is easy to describe broadly and hard to judge without a concrete task, environment, cost and reproducible outcome.
What I’d ask
What does the inference platform expose today? Can a team define an environment, objective, budget and stop condition? Which parts run locally or in a customer account? Are experiments replayable, and can the lab inspect generated data and failed runs? What is the first workflow with a measurable improvement over the team’s current harness?
My editorial take
Keep HABIT on the shortlist for frontier-model research conversations, not as a ready-to-buy world-model platform based on the public pages alone. The next proof is a narrow, replayable experiment with an inspectable artifact and clear compute economics. The direction is interesting; the product boundary still needs to become concrete.
Quick facts
| Field | Sourced detail |
|---|---|
| Product direction | Autonomous research agents for frontier world models |
| Buyer | Model labs, robotics teams and applied-AI researchers |
| Public availability | Homepage links to an inference platform; detailed workflow/pricing not published |
| Founder context | Amazon, Twitch, Radiant Nuclear and JP Morgan experience listed by YC |
| Main question | What repeatable world-model experiment can a customer run today? |
Sources checked
| Source | Checked |
|---|---|
| YC company profile | 2026-09-19 |
| HABIT homepage | 2026-09-19 |
Cohort context
HABIT 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 2026-09-19T16:15:38.947Z 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 ↗