RL Environments Engineer

About the role

ABOUT BESPOKE LABS

Bespoke Labs is an applied AI research lab pioneering data and RL environment curation for training and evaluating agents.

Recently, we curated Open Thoughts https://open-thoughts.ai/, one of the best open reasoning datasets used by multiple frontier labs, trained SOTA specialized models such as Bespoke-MiniChart-7B https://www.bespokelabs.ai/blog/bespoke-minichart-7b and Bespoke-MiniCheck https://www.bespokelabs.ai/bespoke-minicheck, and taught https://www.bespokelabs.ai/blog/improving-multi-turn-tool-use-with-reinforcement-learning agents to do multi-turn tool-calling with reinforcement learning.

Bespoke is uniquely positioned to capture a large market share of data and RL environment curation.

About the Role This is a delivery role. We want an engineer who has built the machinery that turns environment ideas into hundreds or thousands of validated agentic coding tasks, and who can do it here, fast.

You will not be studying environments in the abstract. You will build the pipelines that mass-produce them, design the complex coding worlds agents train inside, and keep pushing throughput: more environments, higher quality, less manual work per task. We will measure you on the volume and quality of environments you ship, not on papers.

The thing we care about most is whether you have done this before. If you have stood up an environment-generation pipeline, scaled agentic task creation into the hundreds or thousands, and shipped it, we want to talk.

WHAT YOU'LL DO

- Build environment-generation pipelines. Own the systems that produce RL environments programmatically, including templating, automated grading, verification, and QA, so the team ships environments at scale instead of one at a time.

- Create complex coding worlds. Build high-fidelity environments around real codebases, with the conventions, dependencies, tooling, and technical debt that real software actually has.

- Scale agentic task creation to thousands. Take task generation from handfuls to hundreds and thousands of validated agentic coding tasks, with automation doing the heavy lifting.

- Build tools that raise throughput. Find the bottlenecks in environment production and remove them. Build the internal tooling and infrastructure that makes everyone on the team faster.

- Own the full task lifecycle. Prompt, environment, grader, running frontier models against the task, failure analysis, and iteration, until each task is rigorous, fair, and hard to game.

- Defend quality at scale. Catch reward hacking and grader loopholes, and build the verification and standards that hold the bar as volume grows.

- Direct coding agents heavily. Use frontier coding agents to build and validate environments faster, judging their output and catching the subtle failures.

Direct frontier coding agents heavily to build and validate environments, judging their output and catching the quiet failures they produce.