Bespoke Labs Raised $40M on July 6. The RL Environment Shortlist Fits on One Page.
Bespoke Labs, Trase, and Sail Research raised $227M in two weeks for RL environments. Here is the sourcing map for a talent pool small enough to name.
On July 6, 2026, Bespoke Labs closed a $40M Series A led by Wing VC to build infrastructure where AI agents can safely learn, test, and improve before production. That is the third RL-environment/eval company to raise real money in roughly two weeks, after Trase ($107M seed) and Sail Research ($80M at $450M). If you are a founder or a recruiter, this window is when "RL environment engineer" stops being a rebrand and becomes a discrete category with a shortlist you can literally name.
What actually got funded
Bespoke's Series A came with a cap table that reads like a talent signal in itself: Wing VC led, Mayfield and The House Fund joined, dbt Labs CEO Tristan Handy wrote a check, and angels from Anthropic, OpenAI, and Meta piled in. The seed was led by 8VC with Jeff Dean, Resolve AI's Spiros Xanthos, and DevRev's Dheeraj Pandey.
The company was founded in 2024 by Mahesh Sathiamoorthy (CEO) and Alex Dimakis (Chief Science Officer). More importantly for sourcing, Bespoke is a core contributor to Terminal-Bench and the team behind OpenThoughts, an open reasoning dataset downloaded more than 500,000 times and used by Thinking Machines Lab, Meta, and Amazon. Every merged PR on those repos is a candidate you can name.
Trase raised its $107M seed on June 24-25, 2026, led by ARCH Venture Partners with Red Cell Partners. The team is 55 people across Seattle and Washington, D.C. President Baskar Sridharan spent about 16 years at Microsoft on Azure storage, then VP eng at Google Cloud, then VP AI/ML services and infrastructure at AWS. Srirama Koneru came from GM of Bedrock Agentic AI Infrastructure at AWS. That is your senior-agent-infra bench: ex-AWS Bedrock Agentic, ex-Google Cloud AI, ex-Azure storage.
Sail Research raised $80M at a $450M valuation. Series A led by Sequoia, seed led by Kleiner Perkins. Founders are Neil Movva (28, ex-Apple, Nvidia, Together AI) and Samir Menon (Stanford, ex-Apple security). Sail's early customers, which double as downstream hiring targets for eval and agent engineers, include Parallel (web search APIs), Detail (code review), Jack & Jill (career agents), and Quadrillion (research agents).
Why this is a new hiring category, not a rebrand
The reflex is to treat "RL environment engineer" as an ML engineer who happens to have read Sutton and Barto. That is the wrong filter and it will give you a bad shortlist.
Environments have to track state faithfully across long-horizon tasks, expose fine-grained telemetry, and emit reward signals that do not accidentally teach agents to cheat. That job is closer to distributed systems, simulation, and game-engine work than to Kaggle-style ML. The person who wrote a robust save/replay system for a multiplayer game is more useful here than the person who tuned a PPO paper on Atari.
There is a second thing the reflex misses. Environments and evals are the same artifact: a dataset, a harness, and scoring rules. Open-source efforts generally treat them as separate. Recruiters searching only for "eval engineer" or only for "RL engineer" each get half the pool. The unified search is the edge, and it is where AI eval infrastructure recruiting starts to look different from generic ML sourcing.
Environments and evals are the same artifact. Anyone hiring for only one of them is sourcing half the pool. </pull> Practically, that means the title "RL Environment Engineer" barely exists on LinkedIn today. The signal lives in GitHub history, not in job titles. If you filter on "PyTorch + RL papers" in an ATS, you will miss the actual builders. ## The repos that are the real resume There are maybe eight repositories that function as the credential system for this category right now. Mine them. - **prime-rl**, **verifiers**, **prime-cli** (PrimeIntellect-ai GitHub org) - **Terminal-Bench** (Bespoke is a core contributor) - **OpenThoughts** (Bespoke's open reasoning dataset) - **GEPA** - **SWE-bench** - **TAU-bench** Sail Research also reports a new BrowseComp-Plus high score of 90.72% at one-tenth the inference cost of rivals, so anyone attached to BrowseComp-Plus contributions or replications is worth a look for agent simulation engineers work. The right query is not "RL engineer in Bay Area." The right query is "people who have merged non-trivial PRs into Terminal-Bench, verifiers, or SWE-bench in the last 12 months, currently at a company that is not Bespoke, Prime Intellect, or Sail." That is exactly the kind of plain-English question no ATS or LinkedIn Recruiter can express, and it is why we built [Refolk](/): you describe the person, and Refolk pulls the ranked shortlist across GitHub, LinkedIn, and the open web in one pass. ## The bear case, and why it makes the hire more urgent Sherwin Wu, OpenAI's Head of Engineering for its API business, said on a recent podcast that he was "short" on RL environment startups. The argument: frontier labs internalize this work, and the third-party market collapses to two or three winners. Take that seriously. Then notice what it implies for hiring. If the market concentrates, the talent concentrates too. The three or four names who ship the winning open-source environment become uncatchable inside 18 months. Anthropic has reportedly discussed spending more than $1 billion on RL environments over the next year, per The Information. That is a demand floor, not a ceiling, and it is why RL environment engineers hiring right now beats hiring in Q2 2027 at any price. METR's benchmarks show the length of tasks AI agents can reliably complete is doubling roughly every seven months. Every doubling is another quarter where the people who can build the harness get more scarce and more expensive.
stat number: 500,000+ label: OpenThoughts dataset downloads note: Bespoke's open reasoning dataset, used by Thinking Machines Lab, Meta, and Amazon. Every listed contributor is a named candidate. </stat>
## The sourcing map, by lane
Six lanes. Work them in this order.
### 1. The three funded companies (and their alumni networks)
Bespoke Labs, Trase, and Sail Research are hiring aggressively. That is obvious. Less obvious: the founders themselves are the map. Mahesh Sathiamoorthy and Alex Dimakis at Bespoke. Neil Movva and Samir Menon at Sail. Baskar Sridharan and Srirama Koneru at Trase. Their prior teams (AWS Bedrock Agentic, Google Cloud AI/ML, Together AI, Apple security, Nvidia) are the concentric circles you work outward from.
### 2. Prime Intellect and Mechanize
Prime Intellect is backed by Andrej Karpathy, Founders Fund, and Menlo Ventures. It launched an "RL environments hub" positioned as a "Hugging Face for RL environments." Will Brown, a researcher there, has been publicly vocal on RL compute cost. Mechanize is the other well-funded specialist in the category. Both are small enough that their engineering rosters are countable.
### 3. Labeling incumbents pivoting to interactive sim
Mercor and Surge are pivoting from static datasets to interactive simulations. Scale AI is doing pieces of this too. The senior ICs who used to run labeling ops at these companies are the closest thing to a trained pool of environment PMs and engineers you will find. This is the most underrated lane on the map, because nobody looking for AI evaluation engineer sourcing thinks to check Surge.
4. Frontier lab eval teams
Anthropic, OpenAI, Meta AI, and Thinking Machines Lab all have internal eval and environments teams. Angels from all three of the first names appeared on Bespoke's cap table, which is a soft signal that the interesting people know each other. These are hard-to-poach but not un-poach-able, especially at the L5/L6 layer where the environments work is less prestigious internally than pretraining.
5. Sail's customer roster
Parallel, Detail, Jack & Jill, and Quadrillion are the four Sail customers named publicly. Each is building an agent product that requires eval and environment expertise in-house. If Sail wins, they are your competition for talent. If Sail loses, they are the acquihire pool.
6. Game engine and simulation refugees
The single most underweighted lane. Faithful state, deterministic replay, telemetry, reward shaping without exploit surface: this is game-engine work. Ex-Unity, ex-Unreal, ex-Roblox, and ex-AAA sim engineers who can code Python are gold. Almost no one is sourcing them for this. When you describe the target as "senior game-sim engineer who has shipped Python since 2023," Refolk will surface people no LinkedIn boolean would ever reach, because the title match is wrong on purpose.
The salary reality
Public reporting notes "some rumor on the street about environment engineers entering the high six figures in salary." Treat that as anecdotal, but directionally correct. When Anthropic is willing to spend nine figures on the category, the individual comp follows.
The investor watchlist for the next raise, which tells you where the next hiring wave lands: Wing VC, ARCH Venture Partners, Sequoia, Kleiner Perkins, 8VC, Mayfield, The House Fund, Founders Fund, Menlo Ventures, Red Cell Partners. When one of them announces the next RL environment round, you already know the sourcing map. When they do, run the same Refolk query against the new company's cap table and open-source footprint and you have a warm list within the hour.
What to do this week
Three moves.
First, pull the contributor lists on Terminal-Bench, OpenThoughts, verifiers, prime-rl, SWE-bench, and TAU-bench. Deduplicate. That is your top-of-funnel and it will be smaller than you expect. Probably a few hundred people globally, of whom fifty or so are the actual signal.
Second, cross-reference against Mercor and Surge engineering rosters. Anyone doing simulation or interactive labeling work there is a warm intro away from being convinced this is the more interesting job.
Third, write the outreach in the language of the work. Not "RL role at a well-funded startup." Something like: "I saw your PR on TAU-bench task X. We're building the harness for agents doing Y and need someone who thinks about reward-hacking the way you clearly do." That is the message that gets an open rate above 40% in this category. Generic recruiting copy gets ignored.
The window is open now because the category is legible now. In six months every recruiter at every frontier lab will be running the same query. Bespoke Labs hiring the right ten people this quarter is worth more than the $40M they just raised. So is yours.
FAQ
How is "RL environment engineer" different from "ML engineer with RL experience"?
An RL environment engineer builds the world the agent learns in, not the agent itself. That means faithful state tracking, deterministic replay, telemetry, and reward functions that resist exploit. The skill set overlaps more with distributed systems, simulation, and game-engine engineering than with model training. Filtering candidates on PyTorch and RL papers will systematically miss the strongest builders, who often come from game-sim or backend infrastructure backgrounds.
Where do I find these people if they do not have the title on LinkedIn?
Almost no one carries the title yet. The credential system is open-source contribution. Mine commit history on Terminal-Bench, OpenThoughts, verifiers, prime-rl, GEPA, SWE-bench, and TAU-bench. Also check the PrimeIntellect-ai GitHub org, the Prime Intellect Discord, and METR's contributor lists. Cross-reference against current employers at Bespoke, Trase, Sail, Prime Intellect, Mechanize, Mercor, Surge, Anthropic, OpenAI, and Thinking Machines Lab.
Is the bear case from OpenAI's Sherwin Wu a reason to not hire here?
The opposite. If the third-party market concentrates to two or three winners as Wu predicts, the talent concentrates with it, and the shortlist gets uncatchable within 18 months. Anthropic has reportedly discussed more than $1 billion on RL environments over the next year, which is a demand floor. Concentration risk on the company side is exactly what makes moving early on the people side rational.
What outreach message actually works for this cohort?
Reference the specific PR, benchmark, or dataset the person contributed to, and describe the reward-hacking or state-tracking problem you are solving in one sentence. Do not lead with the round size or the title. This cohort is engineer-brained and reads generic recruiter copy as noise. A three-sentence note that shows you read their code outperforms a five-paragraph pitch about the company mission every time.