Skip to main content

Open-source research infrastructure · v0.3

Coding agents build better policy systems from environment feedback.

EvoPolicyGym provides a standardized evaluation protocol and a unified interface to interactive environments, giving Coding Agents the infrastructure to evolve executable Policies from bounded feedback and measure them on held-out Cases.

Open leaderboard →Get startedPaper ↗
Package
v0.3.0
Policy boundary
policy/v1
Distributions
63
Named tasks
302
Sol-authored Balatro Policy completing a held-out run with score 1021BalatroSol final Policy · score 1021Agent-authored Crafter Policy navigating a deep-iron combat development episodeCrafter · Deep iron combatDevelopment submission 15 · Episode 10Sol-authored NetHack Policy completing a 1,269-step training episode at dungeon depth 11NetHack · Sol Policy1,269 steps · dungeon depth 11
Three autonomous Policies · real experiment replaysOpen experiments →

Research journal

Recent Blogs

All notes →
Experiment

Into the Dungeon: Building Exploration Systems for NetHack

↗Experiment

Teaching a Coding Agent to write a Balatro policy system

↗Design

Why EvoPolicyGym treats policy programs as research artifacts

↗