Linzwcs/EvoPolicyGym

EvoPolicyGym is infrastructure for evaluating coding agents and generating training experience through Autonomous Policy Evolution.

What it solves

EvoPolicyGym addresses the limitation of standard coding benchmarks that only evaluate a final answer. It provides a framework for evaluating how coding agents can iteratively experiment, learn from feedback, and evolve executable policies (strategies) to solve complex tasks in interactive environments.

How it works

The system operates as a loop between an Agent, a Workspace, and a Server. The Agent edits a program (the policy), submits it for execution across a set of episodes, and receives public feedback. This process is constrained by a fixed "episode budget," forcing the agent to strategically allocate its attempts between exploration and confirmation.

Once the agent finishes, the Host performs a private final selection and assesses the policy on held-out episodes to measure true generalization.

Who it’s for

  • AI Researchers focusing on autonomous policy evolution and reinforcement learning from feedback.
  • Developers of Coding Agents who want to test their agent's ability to improve a strategy over time rather than just writing a single function.
  • Benchmark Authors who can use the provided API to integrate new interactive environments into a standardized evaluation protocol.

Highlights

  • Budgeted Experimentation: Forces agents to manage a limited number of environment interactions to find the optimal policy.
  • Unified Interface: Brings diverse environments (e.g., NetHack, Balatro, Crafter, Gymnasium) under a single benchmark contract.
  • Generalization Testing: Uses private validation and assessment sets to ensure policies aren't just overfit to the training episodes.
  • Agent Integrations: Includes first-party support for CLI-based agents like Claude Code, Codex, and Kimi Code.
  • Extensive Environment Catalog: Supports a wide range of domains including robotics (robosuite, MetaWorld), classic control (CartPole), and complex games (ARC Prize, NetHack).

Related

  • Project
    ServiceNow/BrowserGym

    BrowserGym is an open‑source Python library that turns web browsing into Gymnasium environments, letting researchers plug LLM‑based agents into many web‑automation benchmarks (MiniWoB, WebArena, WorkArena, etc.). It provides a unified API, modular install options, Playwright‑driven browsers, and easy extensibility via an `AbstractBrowserTask` base class. The ecosystem includes AgentLab for large‑scale evaluation and a demo agent that works with OpenAI models.

  • Project
    microsoft/debug-gym

    A text-based interactive debugging framework for Python programs that allows LLM agents to use tools like pdb and bash to investigate and fix bugs in code repositories.

  • Project
    sentient-agi/EvoSkill

    A toolkit for automatically discovering and improving AI skills and prompts for coding agents, evolving general agents into specialists using benchmark-driven feedback.

  • Project
    ANative-Lab/EvoAgentX

    An open-source framework for automatically building, evaluating, and evolving LLM-based agentic workflows through iterative feedback loops.

  • Project
    Farama-Foundation/Gymnasium

    Gymnasium is a community‑maintained Python library that provides a standard, versioned API for reinforcement‑learning environments (classic control, Box2D, Atari, MuJoCo, etc.). It’s a fork of OpenAI’s Gym and serves as the main platform for developing and benchmarking RL algorithms.