autoresearch-agent

Autoresearch Agent

You sleep. The agent experiments. You wake up to results.

Autonomous experiment loop inspired by Karpathy's autoresearch. The agent edits one file, runs a fixed evaluation, keeps improvements, discards failures, and loops indefinitely.

Not one guess — fifty measured attempts, compounding.

Slash Commands

When This Skill Activates

Recognize these patterns from the user:

• "Make this faster / smaller / better"
• "Optimize [file] for [metric]"
• "Improve my [headlines / copy / prompts]"
• "Run experiments overnight"
• "I want to get [metric] from X to Y"
• Any request involving: optimize, benchmark, improve, experiment loop, autoresearch

If the user describes a target file + a way to measure success → this skill applies.

Setup

First Time — Create the Experiment

Run the setup script. The user decides where experiments live:

Project-level (inside repo, git-tracked, shareable with team):

python scripts/setup_experiment.py \
  --domain engineering \
  --name api-speed \
  --target src/api/search.py \
  --eval

User-level (personal, in ~/.autoresearch/):

The --scope flag determines where .autoresearch/ lives:

• project (default) → .autoresearch/ in the repo root. Experiment definitions are git-tracked. Results are gitignored.
• user → ~/.autoresearch/ in the home directory. Everything is personal.

What Setup Creates

results.tsv columns: commit | metric | status | description

• commit — short git hash
• metric — float value or "N/A" for crashes
• status — keep | discard | crash
• description — what changed or why it crashed

Domains

If program.md Already Exists

The user may have written their own program.md. If found in the experiment directory, read it. It overrides the template. Only ask for what's missing.

Agent Protocol

You are the loop. The scripts handle setup and evaluation — you handle the creative work.

Before Starting

1. Read .autoresearch/{domain}/{name}/config.cfg to get:
   • target — the file you edit
   • evaluate_cmd — the command that measures your changes
   • metric — the metric name to look for in eval output
   • metric_direction — "lower" or "higher" is better
   • time_budget_minutes — max time per evaluation
2. Read program.md for strategy, constraints, and what you can/cannot change
3. Read results.tsv for experiment history (columns: commit, metric, status, description)
4. Checkout the experiment branch: git checkout autoresearch/{domain}/{name}

Each Iteration

1. Review results.tsv — what worked? What failed? What hasn't been tried?
2. Decide ONE change to the target file. One variable per experiment.
3. Edit the target file
4. Commit: git add {target} && git commit -m "experiment: {description}"
5. Evaluate: python scripts/run_experiment.py --experiment {domain}/{name} --single
6. Read the output — it prints KEEP, DISCARD, or CRASH with the metric value
7. Go to step 1

What the Script Handles (you don't)

• Running the eval command with timeout
• Parsing the metric from eval output
• Comparing to previous best
• Reverting the commit on failure (git reset --hard HEAD~1)
• Logging the result to results.tsv

Starting an Experiment

Strategy Escalation

• Runs 1-5: Low-hanging fruit (obvious improvements, simple optimizations)
• Runs 6-15: Systematic exploration (vary one parameter at a time)
• Runs 16-30: Structural changes (algorithm swaps, architecture shifts)
• Runs 30+: Radical experiments (completely different approaches)
• If no improvement in 20+ runs: update program.md Strategy section

Self-Improvement

After every 10 experiments, review results.tsv for patterns. Update the Strategy section of program.md with what you learned (e.g., "caching changes consistently improve by 5-10%", "refactoring attempts never improve the metric"). Future iterations benefit from this accumulated knowledge.

Stopping

• Run until interrupted by the user, context limit reached, or goal in program.md is met
• Before stopping: ensure results.tsv is up to date
• On context limit: the next session can resume — results.tsv and git log persist

Rules

• One change per experiment. Don't change 5 things at once. You won't know what worked.
• Simplicity criterion. A small improvement that adds ugly complexity is not worth it. Equal performance with simpler code is a win. Removing code that gets same results is the best outcome.
• Never modify the evaluator. evaluate.py is the ground truth. Modifying it invalidates all comparisons. Hard stop if you catch yourself doing this.
• Timeout. If a run exceeds 2.5× the time budget, kill it and treat as crash.
• Crash handling. If it's a typo or missing import, fix and re-run. If the idea is fundamentally broken, revert, log "crash", move on. 5 consecutive crashes → pause and alert.
• No new dependencies. Only use what's already available in the project.

Evaluators

Ready-to-use evaluation scripts. Copied into the experiment directory during setup with --evaluator.

Free Evaluators (no API cost)

LLM Judge Evaluators (uses your subscription)

LLM judges call the CLI tool the user is already running (Claude, Codex, Gemini). The evaluation prompt is locked inside evaluate.py — the agent cannot modify it. This prevents the agent from gaming its own evaluator.

The user's existing subscription covers the cost:

• Claude Code Max → unlimited Claude calls for evaluation
• Codex CLI (ChatGPT Pro) → unlimited Codex calls
• Gemini CLI (free tier) → free evaluation calls

Custom Evaluators

If no built-in evaluator fits, the user writes their own evaluate.py. Only requirement: it must print metric_name: value to stdout.

Viewing Results

Dashboard Output

Export Formats

• TSV — default, tab-separated (compatible with spreadsheets)
• CSV — comma-separated, with proper quoting
• Markdown — formatted table, readable in GitHub/docs

Proactive Triggers

Flag these without being asked:

• No evaluation command works → Test it before starting the loop. Run once, verify output.
• Target file not in git → git init && git add . && git commit -m 'initial' first.
• Metric direction unclear → Ask: is lower or higher better? Must know before starting.
• Time budget too short → If eval takes longer than budget, every run crashes.
• Agent modifying evaluate.py → Hard stop. This invalidates all comparisons.
• 5 consecutive crashes → Pause the loop. Alert the user. Don't keep burning cycles.
• No improvement in 20+ runs → Suggest changing strategy in program.md or trying a different approach.

Installation

One-liner (any tool)

Multi-tool install

OpenClaw

Related Skills

• self-improving-agent — improves an agent's own memory/rules over time. NOT for structured experiment loops.
• senior-ml-engineer — ML architecture decisions. Complementary — use for initial design, then autoresearch for optimization.
• tdd-guide — test-driven development. Complementary — tests can be the evaluation function.
• skill-security-auditor — audit skills before publishing. NOT for optimization loops.