How to use the Autoresearch agent skill: setup and workflow

What the autoresearch skill makes the agent do

The skill organises an autonomous research project around two loops. The inner loop runs fast experiment iterations, each with a clear measurable outcome, such as making a validation loss go down or testing whether one intervention causes a specific effect. The outer loop periodically steps back to review results, find patterns, update a findings file and decide the next direction.

The agent is told to act as a project manager rather than a domain expert. When a task needs specialised work, it routes to other domain skills in the same repository, for example for paper writing or research ideation.

• Bootstrap: scope the question, search literature, form testable hypotheses and set a baseline metric.
• Inner loop: pick a hypothesis, run the experiment, measure, record, learn, repeat.
• Outer loop: review results, update findings, generate new hypotheses, choose direction.
• Finalize: write the paper through the ml-paper-writing skill, produce a final presentation and archive.

Install the autoresearch skill

Download the ZIP from the skill's page on this directory and unzip it. Then place the folder in one of two locations, depending on whether you want it available everywhere or only in one project.

For all your projects, unzip into ~/.claude/skills/. For a single project, unzip into that project's .claude/skills/ folder. If you use the Claude apps instead, upload the skill under Settings → Capabilities → Skills.

• Personal (all projects): ~/.claude/skills/
• Project only: .claude/skills/ inside your project folder
• Claude apps: Settings → Capabilities → Skills, then upload

Install the extra Python packages for literature search

The skill suggests installing Python packages for literature search. Semantic Scholar and arXiv are both installed with pip, as the skill describes. Exa's web search tool is optional and needs the Exa MCP server to be configured.

• Semantic Scholar and arXiv: pip install semanticscholar and pip install arxiv
• Optional: Exa MCP server, which provides web_search_exa

Trigger the skill with a natural request

You do not need a special command to start. The skill is meant to activate when you start a research project, run autonomous experiments or manage a multi-hypothesis effort. Describe your question in plain language and say you want the agent to run the research.

Be specific about the question and the metric if you have one. If you have a vague idea, the skill says the agent should first have a brief discussion to clarify it. If you already have a plan or a research-state.yaml file, say so and the agent will resume from that state.

• "Start an autoresearch project on whether sparse attention preserves accuracy on long documents. Use validation loss as the metric."
• "Resume the autoresearch project in this folder and continue from research-state.yaml."
• "I have a research proposal in proposal.md. Review it, set up the workspace and start the experiment loops."

Keep the agent running with the wall-clock loop

The skill's reference on agent continuity says the wall-clock loop should be set up first. Without it, the agent may complete one cycle and stop. On Claude Code, the skill uses the /loop command with a 20-minute interval. This loop is separate from the research loops, so the two do not need to line up. OpenClaw can use a heartbeat for continuous operation instead.

What the project folder contains

The skill creates a project folder with a central state file, a decision log and a findings file. Each hypothesis gets its own experiments folder with a protocol, code, results and analysis. Templates for the state file, log, findings and progress presentation are included in the skill's templates folder.

Large files such as model checkpoints should go to a separate storage location, not the project directory.

• research-state.yaml: central state tracking
• research-log.md: decision timeline
• findings.md: evolving narrative synthesis
• literature/, src/, data/, experiments/, to_human/, paper/

Safety and what to check before running it

The skill runs scripts and uses the network. It tells the agent to act without asking for confirmation, so the agent may run code or use resources while you are away. Only run it where that is acceptable. Do not point it at production systems, paid cloud compute or real user data unless you have checked the setup yourself.

Treat the agent's literature summaries and conclusions as drafts. The skill's own guidance is that the agent may be wrong and results need human review.

• Runs scripts and makes network calls.
• Works without asking permission at each step.
• Keep experiments away from production systems, paid compute and real user data unless you have reviewed the setup.
• Review findings and presentations before relying on any result.