Who it is forA researcher or developer who wants to see how an action preference changes a self-evolving harness, without an API key or Docker.
The problemReading a self-evolving harness by hand is slow, and it is hard to tell whether a change came from the preference or from noise.
The outcomeA per-arm table of notes and tools built, a behavioural R and p value, and a local visual replay of the runs.
What you need first
- Python 3.10 or newer
- pip to install proteus-evolve==0.4.0 (no model SDK is required for the minimal harness)
- For real harnesses: Docker and a DeepSeek or Codex key, as listed in the harness table
Steps
pip install proteus-evolve==0.4.0proteus run --harness minimal --arm neutral --arm review:notes --arm review:tools --seeds 4 --episodes 8 --out runs/<target>(inferred, check the README)proteus measure --harness minimal --out runs/<target>(inferred, check the README)proteus visualize --out runs/<target>then open http://127.0.0.1:8301 in a browser (inferred, check the README)
Prompt for your agent
In the Proteus repo (proteus-evolve/Proteus), install proteus-evolve==0.4.0 and run the minimal harness with the neutral, review:notes and review:tools arms, writing output to runs/<target>. Expected output: a per-arm table of notes and tools counts plus a behavioural R and p value from proteus measure. Verify that the measure output includes a p value and that the review arms show a different notes or tools count than neutral.
Caveat. The minimal harness uses a mock policy, so the numbers show the pipeline working offline, not how a real model behaves.
Use this in your AI agent
Ask your agent
Using the Deduped MCP, get the setup steps for proteus-evolve/Proteus and "Compare evolution arms offline in one afternoon"
Add Deduped to Claude Code once
claude mcp add --transport http deduped https://dedupedai.com/mcp
Every Monday: the 5 hottest repos and exactly how to use them. Free, one email a week, unsubscribe anytime.
Get them freeShare: X LinkedIn Hacker News Email Copy link
Third-party repos we don't own. AI-written use cases can be wrong. Disclaimer