OpenEnv documentation
Coding Agent Training with TRL (Pi)
Coding Agent Training with TRL (Pi)
This tutorial covers the black-box training path: training the actual pi coding agent, with its own planner,
tools, context management, and stop condition, using TRL’s experimental AsyncGRPOTrainer. The agent owns its loop, and OpenEnv captures what it did.
Three GRPO patterns, three tutorials. For a standard
reset()/step()flow where TRL drives the episode, see the Wordle GRPO tutorial. For harness rollouts where the trainer still generates each turn (white-box), see the BrowserGym harness tutorial. Use this page when you want to train a production agent as-is, without reimplementing its loop.
How It Works
The full recipe lives in TRL. The moving pieces:
- Each rollout runs the agent inside an OpenEnv session created by
PiSessionFactoryfrompi_env, intransparent_proxymode. A small proxy inside the sandbox forwards the agent’s/v1/chat/completionscalls to your vLLM server and records each turn’s token ids and logprobs to a trace. - When the agent stops, TRL’s
HarnessRolloutWorkerreads the trace, rebuilds the per-turn training rows from the recorded ids, and scores the final workspace with the session’sverify()method (a held-out verifier the agent never sees). AsyncGRPOTrainertrains on those rows, propagating the rollout reward to every trained token through the group-relative advantage. NCCL weight sync keeps the vLLM server on the current policy, so the agent always samples from the model being trained.
Each rollout gets its own isolated session: one sandbox, one proxy port, one
agent process. Three small functions adapt the recipe to your task: rollout_reward_fn (outcome to scalar reward), train_turn_fn (which turns
receive gradient), and agent_turn_fn (which trace entries are real agent
turns rather than auxiliary calls like title generation).
Full Recipe
The reference script trains on competitive-coding problems from agentica-org/DeepCoder-Preview-Dataset: the agent writes solution.py, and
the verifier runs it against held-out tests, returning the fraction passed. It
is self-contained, runs the agent in a local subprocess sandbox (no container
setup needed), needs two GPUs (one serving the policy with vLLM, one
training).
examples/scripts/openenv/pi.pyin TRL: the complete, runnable script.envs/pi_env: the OpenEnv side, including the session factory, sandbox backends, and the transparent interception proxy.