RL environment jobs: building environments for AI agents

What RL environment jobs are

RL environment jobs pay software engineers to build the tests AI agents are trained and graded on. In reinforcement learning, a model attempts a task, something checks whether it succeeded, and the result feeds back into training. The environment is the task plus everything needed to run and check it. micro1's senior backend listing puts it neatly: you build the exam, not sit it.

This is the most engineering-heavy corner of AI training work. For how it relates to labeling and evaluation, see what AI training work actually is.

What one environment contains

The listings agree on the parts. Invisible's Real-World Terminal and Infrastructure Scenarios listing names four:

1. A problem setup drawn from real work: a defect to fix, a feature to add, a refactor, a slow path to speed up, or a broken piece of infrastructure.
2. A reproducible environment the agent runs in, which behaves the same way every time it is rebuilt. That means pinned dependencies and deterministic builds.
3. A reference solution, often called the golden solution, proving the task can be done.
4. An automated verifier that decides whether an attempt passed.

The verifier is where the effort goes. Too strict and it rejects a correct solution written differently from yours; too loose and a wrong one passes. micro1's backend environments add deliberately broken variants, so the model has to diagnose rather than pattern-match, and its competitive programming track asks for original problems that do not exist on any online judge.

Mercor's Go Codebase Q&A track is a variation: tasks arrive pinned to a commit in a real open-source Go project, and you rewrite the question until frontier coding agents fail it while it stays fair, adjusting rubrics and plausible wrong answers, then run a validation loop until every stage passes.

Who hires for it

Of the 699 listings on this site on 29 September 2026, 14 are for building RL environments or coding tasks with verifiers. micro1 posted 12, Mercor 1 and Invisible Technologies 1. Five more are auditor roles that review this kind of task rather than write it: four at Mercor and one at Invisible.

Ten of the 14 are pitched at senior level and the other four at mid level. The languages named most often are Python, Java, Rust, Go, C++ and TypeScript, and some tracks specialise: MCP tool use, cloud infrastructure, security, competitive programming.

What it pays

• micro1 contract roles: eleven hourly bands from $30 to $150. Its two RL Environments postings carry identical text at $50–100 and $100–150.
• Mercor Go Codebase Q&A: $130 per task.
• Invisible terminal scenarios: a flat $35 an hour.
• Auditors: $70–90 an hour on Mercor's four, $60 on Invisible's.
• Salaried: micro1's Member of Technical Staff, Research Engineering lists $200,000–300,000 a year for architecting environments, reward functions and verifiers.

The micro1 contract listings we read state that pay is output-based: per task that meets the specification, with a weekly minimum number of submissions the ads leave unquantified. An environment that fails review is unpaid time, and a hard verifier can take far longer than planned the first time. Ask what a typical task takes before you commit; our payout guide explains why.

Browse current roles under Software & IT.