a real run, replaying · y = tool calls per episode · click a dot for the episode · click empty space to zoom out · drag the bar to scrub

acting on

pause · 0.5× 1× 2× 4×

an open instrument · est. 2026 · MIT

Self-evolution for any agent harness.

Plug in. Evolve. Measure.

Enter the evolving lab Star on GitHub

↓

the specimen · research fleet · control arm — no goal, no injected preference

click any dot to step inside an episode · scroll on for the instrument

The instrument

Any harness

Plug in DeepSeek Harness, Pi, our reference harnesses, or your own — one small adapter, no harness code modified. The framework, sandbox, and measurement work unchanged.

Goal or no goal

Run evolution under one goal, many goals, or none at all — with evaluators the agent sees, or never sees. No-goal, unpressured evolution is a first-class mode.

Measured, not just scored

Structural distance per surface, behavioural distance with a permutation test, and a crystallization test: remove the injected preference and read back what the harness built on its own.

The evolve loop

context-fresh each episode · only files cross the boundary · print version

episode boundary: git snapshot · evaluators run · accept/reject — every run keeps its full history as a git repository

60-second demo · no key · no docker

pip install proteus-evolve   # or: git clone github.com/proteus-evolve/Proteus && pip install -e .
proteus run --harness minimal --arm neutral --arm review:notes --arm review:tools \
    --seeds 4 --episodes 8 --out runs/demo
proteus measure --harness minimal --out runs/demo