Skip to content
All projects

RL environments

Give agents real work. Measure what they can do.

My work
Former Member of Technical Staff at Proximal Labs
Status
Previously at Proximal Labs
Built with
Environment design, Agent evaluation, Benchmark infrastructure
Environment loopProximal Labs
Make progress observable.

What an environment does

A reinforcement learning environment gives an agent a task, tools, a world it can change, and feedback on the result. For engineering work, that means making the repository, execution conditions, and scoring behave like a coherent system.

My work

At Proximal Labs I built environments for frontier AI research, along with agents and techniques that generate those environments. My work also included investigating how coding agents plan, use tools, and fail on difficult tasks.

A public example: FrontierSWE

I contribute to FrontierSWE, a public benchmark for long-running engineering tasks. Its first release includes 17 tasks across implementation, performance, and research, with a 20-hour budget per task. The public announcement explains the benchmark design and results.

Why this is interesting

A useful task needs more than a plausible prompt. It needs a reproducible starting state, meaningful constraints, and a way to distinguish a correct result from an answer that merely looks finished. That combination makes environment engineering a systems problem.

Read the FrontierSWE announcement

The build, in writing.

1 notes on the architecture, experiments, and tradeoffs.

Read the series
  1. An RL environment is a systems problemA task, a sandbox, and a score need to agree. The engineering around that contract is where things get interesting.3 min