hgym — Harness Gym

Environments for agents whose policy boundary is the tool surface.

hgym is a fresh start of LLM Gym, rebuilt around one idea: the tool surface — the set of tools an agent can call — is the program. Same environment, same model: {search, compose, terminate} is one research program; {search, compose, draft, critique, revise, terminate} is draft-and-revise; {answer, terminate} is one-shot. Surfaces compose as config diffs, not Python forks, which makes them a target an optimizer can range over.

Environments are ToolUsingEnvs: a task loader, an initial observation, a pure verifier, and a set of MCP servers. Tools are MCP servers (in-process, stdio, or HTTP); episodes are session-keyed for safe concurrent rollouts; termination is a reserved terminate tool or the horizon; verification is a pure function over the recorded trajectory.

Design goals:

• Zero infrastructure. pip install hgym, export an API key, run an episode. No Docker, no databases, no gateway servers. Traces are local JSONL files.
• Provider-neutral. A thin model-client seam speaking the OpenAI-compatible wire schema; bring any provider, proxy, or local server.
• The optimizer workflow is first-class. Export an environment's harness as an editable directory (templates, model, tool manifest), let a human or an agent edit it, re-run, score.

Status

Pre-alpha. The core engine is ported and under active development; see the spec and roadmap. Not yet ready for use.

Quickstart (target API)

import hgym

rollouts = await hgym.run_episodes(
    env_name="wordle_v1",
    model="openai/gpt-5.2-mini",
    num_tasks=50,
)

License

Apache-2.0. Portions derived from llmgym (© TensorZero, Apache-2.0) — see NOTICE.