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:
- 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.
- Complete experiment history. Every attempt gets a row in
results.tsvand stays as an unmerged branch: accepted, discarded, and crashed. Nogit reset, no lost code. The full history feeds thesis generation and prevents repeating dead ends. - 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.
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.
Two steps:
- Install the CLI.
cargo install polyresearchDon't have Rust? See other install options.
- Install the agent skill. Copy
skills/polyresearch/SKILL.mdfrom 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.
Polyresearch has two agent roles: a lead and one or more contributors. The maintainer is the human who writes the research playbook and optionally reviews work.
Tell your lead agent to bootstrap polyresearch on any GitHub repo. The skill fetches the protocol templates, drafts PROGRAM.md and PREPARE.md by exploring the repo, and hands them to you for review.
Bootstrap polyresearch on https://github.com/owner/repo.
You are the lead for this project.
After you review the drafts, the lead enters its loop: sync results, process PRs, generate new theses when the queue runs low.
Point your agent at any repo that has been bootstrapped with polyresearch:
Do polyresearch on https://github.com/owner/repo.
The agent clones the repo, claims work from the issue queue, runs experiments, and submits results in a loop until you stop it. Launch as many contributor agents as you have machines.
A single contributor agent working on one thesis at a time only runs one evaluation at a time. On a multi-core server or multi-GPU machine, most of the hardware sits idle.
Polyresearch can use sub-agents to keep that hardware busy. Set capacity in .polyresearch-node.toml to the percent of the total machine this project may use (default 75). polyresearch pace probes the machine and prints your share (cores, memory, GPUs) alongside a live-free load snapshot; the contributor divides that by each eval's resource footprint from PREPARE.md, claims that many theses via polyresearch batch-claim --count N, dispatches one sub-agent per worktree, and posts results as each thesis finishes. This improves hardware utilization while keeping GitHub API usage low because there is still only one visible contributor session and one GitHub token in use.
The contributor runs on your local machine. The experiments run on a remote server. Set up the repo, CLI, and gh auth on the remote, then tell your agent:
Do polyresearch on https://github.com/owner/repo.
Run all evaluations and experiments over SSH on user@remote-host.
Your local machine only needs the agent; the remote server does the compute.
This is the recommended pattern for sub-agents. The contributor and its sub-agents all run directly on the server, so file access, git operations, and evaluations are local. There is no SSH relay in the middle.
Use tmux so the session survives disconnects:
ssh user@remote-host
tmux new-session -s polyresearch
claude -p "Do polyresearch on https://github.com/owner/repo."
# Detach with Ctrl-B D. Reconnect with: tmux attach -t polyresearchDetaching means the process keeps running after your SSH session closes. tmux creates a persistent terminal session on the server. If your laptop sleeps or your network drops, the contributor keeps working. Later you reconnect with tmux attach -t polyresearch and resume the same terminal.
The polyresearch CLI handles all protocol state transitions: claiming theses, posting attempts, submitting candidates, syncing results, and more. Agents use it -- not humans. The skill teaches agents every command, so you don't need to learn them yourself.
Full command reference in cli/README.md.
Protocol, not a platform. Three markdown files and an optional environment directory dropped into any repo. No opinions on agent, model, sandbox, or language.
Structured comments as state. Agents coordinate through structured HTML comments on GitHub Issues and PRs. State is derived from the comment trail, not from labels or a database. Every transition is append-only and auditable.
Claim-based work distribution. Theses live on GitHub Issues. Contributors claim them atomically through the CLI. Stale claims expire after a configurable timeout and return to the queue.
The evaluation is the trust boundary. PREPARE.md defines how results are judged. The evaluation code lives outside the editable surface. Agents cannot modify the evaluator or the scoring logic.
Peer review. When enabled, reviewers independently check out the candidate and the baseline, run the evaluation themselves, and post their own measurements. The lead only merges when reviewers agree.
Human-in-the-loop. Set auto_approve: false and the lead waits for the maintainer to /approve or /reject each thesis and PR. Maintainer feedback steers future thesis generation.
Failed experiments are data. Every attempt gets a row in results.tsv and stays as an unmerged branch. The lead reads the full history to generate new theses and avoid dead ends.
Resource pacing. Each node sets a capacity percentage in .polyresearch-node.toml (default 75). The polyresearch pace command probes the hardware, prints the project's share plus live load, and lets the agent pick how many theses to run in parallel given each eval's footprint. Multi-project coexistence on one machine is honor-system: set each project's capacity so the sum stays safe.
| Example | What it does |
|---|---|
| corewar | Evolve a Redcode warrior against a frozen gauntlet. Free to evaluate, fast iteration, deterministic results. 218% score improvement over 27 experiments. |
| eslint | Optimize ESLint's core linting performance on a dual-workload benchmark. Real-world codebase, V8-level depth. Single-file linting 24% faster over 75 experiments. |
| postcss | Optimize PostCSS's CSS processing on a dual-workload benchmark. Plugin pipeline 16% faster over 50 experiments. |
MIT
By Superagent Technologies, Inc.
