May 22, 2026 · Guide
RL Environments for Coding Agents
Coding is the most commercially active RL environment domain. Here's how code environments work, what makes a good one, and the companies building them in 2026.
Code is the single most active domain in RL environments, and for a good reason: software engineering is one of the few real-world agent tasks where the reward signal is essentially free. If the tests pass, the agent succeeded. If they don't, it failed. No human grader, no LLM judge, no ambiguity. That clean signal makes coding the easiest domain to scale RL training in, which is why most of the well-known environment specialists started here.
What a code environment actually contains
A useful code environment is more than a sandbox with a compiler. It needs: a real repository (not a toy snippet), a real task expressed as a failing test or a bug report, an isolated execution layer that resets cleanly between attempts, and a grader that runs the tests and returns a structured pass/fail signal — ideally with partial credit when only some tests pass. The leading vendors also instrument the agent's trajectory: which files it opened, what edits it made, how many tool calls it used. That trajectory data is often more valuable than the final score.
Depth versus breadth
Two strategies have emerged. Some vendors build a small number of extremely high-fidelity environments — real codebases, real bugs, hand-crafted graders. Mechanize is the clearest example: a tight set of software-engineering environments and evals built for frontier coding agents, with graders that score performance on actual SWE tasks. The bet is that depth beats breadth for training the very best agents.
Others build many smaller, more synthetic environments to support fast iteration loops. Datacurve leans this way, pairing code-execution environments with high-quality coding data for tight training loops where you want millions of attempts more than you want realism.
A third group sits in between, building benchmark-style code environments framed around realistic developer workflows rather than toy problems — AfterQuery does this for both code and finance.
What to look for as a buyer
Match the vendor strategy to your training stage. Pre-training and early post-training benefit from breadth — lots of tasks, lots of variety, cheap rollouts. Late-stage RL on a flagship coding model benefits from depth — fewer environments, but each one a faithful replica of the kind of work a senior engineer actually does. Most serious training programs end up buying both.
Browse the rest
See every code-environment company in one place, filterable by domain, on the RL environment companies list.