Claude Skill MartBrowse skillsWhat is a Skill?
 
← Back to list

Autoresearch Agent (Autonomous Experiment Loop)

An autonomous edit–measure–keep loop that optimizes a single file against any measurable metric.

AutomationAdvanced★ 24,151⑂ 3,405AI score 8/10Last updated: Aug 9, 2026

What it does

An autonomous optimization loop inspired by Karpathy's autoresearch. Give it (1) a target file, (2) an evaluation command that prints a metric, and (3) a git repo — then the agent loops indefinitely:

  1. Read results.tsv history to see what worked and what hasn't been tried
  2. Make exactly one change to the target file and commit it
  3. Run the evaluation script and read the metric
  4. Better → KEEP the commit; worse → DISCARD via git reset --hard; error → log CRASH

It ships with /ar:setup, /ar:run, /ar:loop, /ar:status, /ar:resume commands, free evaluators (speed, size, memory, test pass rate, build time) and LLM-judge evaluators (headlines, prompts, copy). Every 10 runs the agent reviews patterns and updates the Strategy section of program.md so future sessions inherit what it learned.

Who it's for

  • Backend/performance engineers who spend days chasing benchmark numbers
  • Platform engineers shaving bundle size, build time, or Docker image weight
  • Marketers and prompt engineers who want dozens of scored variants of a headline, ad, or system prompt
  • Anyone who wants to leave an agent running overnight and read a dashboard in the morning

Example uses

  • Cut API latency: target src/api/search.py, metric p50_ms from pytest bench.py, 47 runs down to 185ms
  • Slim a frontend bundle: use the benchmark_size evaluator on size_bytes to iteratively shrink build output
  • Raise headline CTR: target content/titles.md with the llm_judge_content evaluator scoring 0–10, letting the agent search for higher-scoring titles

· · · Install guide · · ·

Install in the Claude app (no terminal)
  1. Download the ZIP with the button below.
  2. In Claude, open Settings → Capabilities and turn on 'Code execution and file creation'. (one time)
  3. Go to Customize → Skills → + → 'Upload a skill' and upload the ZIP.
↓ Download ZIP
Install in Claude Code

Let Claude do it — paste this into Claude Code

Install the skill I found on Claude Skill Mart.
Copy the .gemini/skills/autoresearch-agent folder from the GitHub repo alirezarezvani/claude-skills into my ~/.claude/skills/autoresearch-agent/.
When it's done, tell me in one line what this skill can do.

Install with a command instead

git clone https://github.com/alirezarezvani/claude-skills.git && mkdir -p ~/.claude/skills && cp -r claude-skills/.gemini/skills/autoresearch-agent ~/.claude/skills/

⚠ This is a third-party skill. Check the source repository before installing.

  1. Open a terminal and clone the repo: git clone https://github.com/alirezarezvani/claude-skills.git
  2. Copy the skill into Claude Code: mkdir -p ~/.claude/skills && cp -r claude-skills/.gemini/skills/autoresearch-agent ~/.claude/skills/
  3. Start Claude Code inside the project you want to optimize. It must be a git repo (git init plus an initial commit).
  4. Commit or stash any work in progress, then create a dedicated branch: git checkout -b autoresearch/engineering/api-speed
  5. Ask Claude to "set up an autoresearch experiment to optimize <file> for <metric>", or run /ar:setup.
  6. Verify your evaluation command prints metric_name: value by running the --dry-run mode first.
  7. Run /ar:run for a single iteration, then start /ar:loop for autonomous mode and check progress with /ar:status.
View source on GitHub ↗License: MIT