polyresearch

Distributed autoresearch. Multiple machines, multiple contributors, verified results.

Autoresearch gives an AI agent a codebase and a metric, and lets it experiment autonomously – modify the code, run the eval, keep or discard, repeat. You wake up to a log of experiments and a better result. But it runs one agent on one machine. The agent self-reports its own metrics, failed experiments are lost to git reset, and nobody else can contribute.

Polyresearch keeps the same loop and adds three things:

1. Open participation. Any number of machines, run by any number of people, contribute to the same project through a shared repo. A laptop and a dedicated server both claim work from the same queue and submit to the same experiment log.
2. Complete experiment history. Every attempt gets a row in results.tsv and stays as an unmerged branch: accepted, discarded, and crashed. No git reset, no lost code. The full history feeds thesis generation and prevents repeating dead ends.
3. Independent verification. Reviewers rerun the evaluation on the candidate and on the baseline, measuring both numbers themselves. The evaluation code lives outside the editable surface, so agents cannot grade their own homework.

How it works

A polyresearch project is any GitHub repo with a few coordination files:

• PROGRAM.md — the research playbook. Same concept as autoresearch’s program.md. Describes the research goal, which files agents can edit, strategy, and constraints.
• PREPARE.md — the evaluation setup. What commands to run, how to parse the metric, what the ground truth is. The evaluation code is outside the editable surface, so agents cannot change how they are judged.
• POLYRESEARCH.md — the coordination protocol. Same for every project, like a LICENSE file. Not modified.
• .polyresearch/ — the reproducible environment. Setup scripts, evaluators, frozen dependencies. Optional.

Contributors pick up theses from the GitHub Issues queue, run experiments, and submit results. Other contributors independently verify results. The lead manages the queue and merges accepted work. Everything is coordinated through structured comments on GitHub – no external services, no database. Requires git and gh.

Install

Two steps:

1. Install the CLI.

cargo install polyresearch

Don’t have Rust? See other install options.

2. Install the agent skill. Copy skills/polyresearch/SKILL.md from this repo into your agent’s skill directory (e.g. ~/.claude/skills/polyresearch/, or equivalent for your agent). The skill teaches agents the full protocol – bootstrapping, the lead loop, the contributor loop, and all CLI usage.

Usage