If you run a software company and want to sell RL environments or expert coding data to AI labs, the first thing you learn is that the biggest buyers of expert work do not buy from companies. They buy from people, one account at a time. That is not a rumour. It is written in their terms, and the terms are worth reading before you plan a sales motion around a platform that will never let a team in. Below is what each platform says, as of October 2026, and the routes that do work for an agency with its own engineers.
What the platforms say, in their own words
Mercor
Mercor's worker terms (updated August 26, 2026) open with "This is a contract between you and Mercor," and forbid sharing "your password(s) and/or any other authentication credentials with anyone else." The tax and work authorisation policy is the decisive line for a company: "Complete your tax form (W9 or W8-BEN) as an individual," and "We are unable to accept W8-BEN-E or other forms related to business entities." The support page adds that "The agreement is standardised for all contractors." No programme for supplying a team of workers is published on mercor.com; the one company-facing seller route there is data licensing (mercor.com/data), and the partner page is for buyers.
Outlier (Scale AI)
Outlier's terms of use (effective May 11, 2026) state: "You may only possess one Account," "you may not allow third parties to use your Systems Accounts," and "you may not assign or otherwise transfer your Services accounts to any other person or entity." Payment accounts must be "in your own name." A company cannot hold an Outlier account on behalf of its engineers.
Handshake AI
Handshake's fellow support page (updated October 20, 2026) says "All Fellows are hired as 1099 independent contractors" and "Fellows are contracted and paid directly by Handshake, not the partnering AI lab." The programme page requires that "you must be based in the U.S. and have a valid work authorization" that Handshake can support, and the original work policy requires that "Your contributions must be original and created solely by you." Three separate walls for a company with engineers abroad.
Micro1
Micro1's candidate terms (updated July 21, 2026) require that you "Not share your account credentials or transfer your account to any other individual," and that interview answers "will be your original work product." The company's contact form is "for companies and clients," meaning buyers, and its Data partnerships programme buys operational data from established companies, not engineering labour.
Surge AI and Turing
Surge AI publishes no contributor-specific terms; its public terms of service date from 2021 and its RL environments architect role is an internal hire, not a vendor programme. Turing describes its talent route as joining an expert network. Neither publishes a way for a company to supply a team. We treat both as closed to agencies until they say otherwise.
The pattern is consistent, and it is not hostility to agencies. These platforms sell labs an individual-level guarantee: a named person, verified identity, one account, original work. A team routed through one login breaks the guarantee, so the terms forbid it. The right response is not to work around the terms. It is to sell where companies are the expected seller.
Route 1: codebase and environment marketplaces
DataVendor, run by HUD, was "built for sellers with code, data, and the skills to create training tasks," and lists "prepared code repositories, raw data ZIPs, and HUD tasksets with their reinforcement learning (RL) environments." HUD's own guide to selling a codebase (June 24, 2026) is explicit about who it is for: "Founders and consultancies with a serious modern codebase." It also states the economics: pricing "starts at a $5,000 baseline per codebase and increases with grading," and "The seller receives 80% of the final payout, with 20% covering grading and packaging." This is the route where a company that owns a production codebase, or has the rights to one, can list it and the tasks built on it under the company's own name.
What it needs from you: clean rights to the code, a scrubbed snapshot, and tasks that survive grading. The grading step is where most of the value and most of the rejection sits, which is why we built a validator before we built a listing.
Route 2: open environment hubs with team accounts
Prime Intellect's Environments Hub accepts uploads from individuals and organisations: the documentation shows a push under a user account and, with one flag, "to a team." Its environments programme (October 27, 2025) committed "hundreds of thousands of $ as grants," with open bounties "in the $100-500 range" and application-only bounties "ranging from $1000 to $5000+." The money per environment is small next to lab contracts, but the route is open, public and company-friendly, and a published environment is a portfolio piece a lab buyer can run.
Route 3: sell to the labs and the startups that supply them
The supply side of this market is companies, not freelancers. RL List (updated October 5, 2026) is "a cited directory of the 42 companies building RL environments for frontier AI labs"; RL Atlas lists 86. Epoch AI's FAQ on RL environments (January 12, 2026), built from 18 interviews, describes contract sizes "often six to seven figures per quarter" and notes that "Traditional data providers like Mercor, Surge, Handshake, and Turing" now sell environments too. Wing VC counted "roughly 20 seed- to Series A-stage companies on relatively similar footing" in January 2026. TechCrunch reported on September 16, 2025, citing The Information, that "leaders at Anthropic have discussed spending more than $1 billion on RL environments over the next year."
For an agency this means two doors. The first is the labs themselves, which contract with vendors directly and are building vendor-facing tooling (a since-removed Anthropic posting for a full-stack engineer in reinforcement learning described building "vendor-facing interfaces and tooling that let external partners create, submit, and iterate on training environments"). The second is the environment startups and data providers above, who take environment orders larger than their own teams can fill. A pod that delivers graded tasks in Harbor and HUD formats under one contract is what they subcontract.
What to bring to the door
Whichever route, the buyer will ask the same things, and the answer cannot be a deck:
- Rights. Who owns the codebase, who owns the tasks, and the chain of IP assignment from each engineer to the company to the buyer.
- Identity. Which named engineers touch the work, verified for identity and location, with project-only access. The platforms demand this per person; a company must demand it of itself.
- Grading evidence. Validator output per task: base runs, gold runs, empty run, a reward-hacking note and a calibration row. The marketplace grades you on it; the lab will too.
- A pilot shape. Fixed price, a defined task count, delivered in the buyer's format and graded by the buyer with their own models. Nobody signs a seven-figure quarter with a vendor they have not tested.
We built the RL environments line around exactly those four items, and the hire RL engineers page describes the roles behind them. If you are an agency weighing this market: do not spend a quarter trying to get a team through a contractor platform. Spend it getting one environment through grading.
Last updated October 26, 2026