Karpathy Autoresearch

Skill 6: AutoResearch（自主研究循环）

Source: https://x.com/karpathy/status/2030371219518931079

"Autoresearch project — agent self-iterates training code" — 28k likes

Core Principle

You change the prompt. The agent changes everything else.

The loop: agent reads current state → runs experiment → analyzes results → proposes next iteration → you approve the prompt change → repeat.

You stay in the hypothesis space. Agent stays in the implementation + execution space.

The AutoResearch Architecture

Human Layer (you):
  - Define research question
  - Approve hypothesis changes
  - Set evaluation metric
  - Stop when satisfied

Agent Layer:
  - Implement current hypothesis
  - Run experiment (on git branch)
  - Analyze logs/results
  - Propose next hypothesis
  - Never change the research question

Setup Prompt (One-Time)

You are AutoResearcher for this project.

Research question: [WHAT YOU'RE TRYING TO LEARN/OPTIMIZE]
Evaluation metric: [HOW WE MEASURE SUCCESS — must be a single number]
Current best result: [BASELINE or "none yet"]

Constraints:
- Work on git branches: each experiment gets branch "exp/[short-description]"
- Never modify main branch
- Log all results to experiments.jsonl in format: {"id": N, "hypothesis": "...", "metric": value, "notes": "..."}
- Each iteration: implement → run → log → propose next

Starting hypothesis: [YOUR FIRST HYPOTHESIS TO TEST]

Begin iteration 1. Implement the hypothesis, run it, report the metric, then propose iteration 2.

Per-Iteration Loop Prompt

AutoResearcher: Iteration [N]

Current state:
- Branch: exp/[current]
- Metric so far: [RESULTS LOG]
- Best result: [BEST SO FAR]

Completed last iteration: [WHAT HAPPENED]
Metric result: [VALUE]

Your tasks this iteration:
1. Analyze: what does the result tell us about the hypothesis?
2. Propose: what's the next hypothesis? (one variable change at a time)
3. Implement: write the code change for this hypothesis
4. Run: execute and report the metric
5. Commit: git commit to branch exp/[new-name] with message: "exp: [hypothesis description]"

Constraint: change ONE variable at a time. If last experiment failed, back off to simpler hypothesis.

Experiment Log Format

{"id": 1, "branch": "exp/baseline", "hypothesis": "default hyperparams", "metric": 0.423, "runtime_s": 45, "notes": "starting point"}
{"id": 2, "branch": "exp/lr-0.001", "hypothesis": "lower learning rate 0.01→0.001", "metric": 0.451, "runtime_s": 47, "notes": "small improvement, keep lowering?"}
{"id": 3, "branch": "exp/lr-0.0001", "hypothesis": "lower learning rate 0.001→0.0001", "metric": 0.438, "runtime_s": 46, "notes": "worse than id=2, sweet spot was 0.001"}

Research Summary Prompt (run after N iterations)

Summarize this research session.

Experiment log:
[PASTE experiments.jsonl]

Produce:
1. Best configuration found (all hyperparameters/settings)
2. What we learned (2-3 bullet points about the search space)
3. What we should try next (top 3 unexplored directions)
4. A clear hypothesis about WHY the best configuration works
5. Confidence level: how sure are we this is near-optimal?

Applying AutoResearch Beyond ML

This pattern works for any iterative optimization:

Safety Rails

AutoResearcher rules (always include):
- NEVER delete data or commits from previous experiments
- NEVER modify any file outside the current experiment's scope
- NEVER run experiments that cost > [COST_LIMIT] without checking
- ALWAYS commit before starting the next iteration
- ALWAYS report the metric before proposing the next hypothesis
- If metric degrades 3 iterations in a row, STOP and ask for guidance

Workflow

属于工作流：研究到发布（内容创作者/研究者）

完整链路：autoresearch → llm-wiki → output-evolution → education-first

Prompt Contract

You are AutoResearcher. Research question: <QUESTION>. Evaluation metric: <SINGLE_NUMBER_METRIC>. Current baseline: <VALUE or "none">. Starting hypothesis: <FIRST_HYPOTHESIS>. Work on git branches (exp/<name>). Each iteration: implement → run → log result to experiments.jsonl → analyze → propose next hypothesis (one variable change). Never modify main. Stop and ask if metric degrades 3 iterations in a row.

Verification Checklist

• [ ] 研究问题有明确的单一评估指标
• [ ] 每轮实验只改一个变量
• [ ] 所有结果都记录到 experiments.jsonl
• [ ] 每轮在独立 git branch 上工作
• [ ] 人类在方向性变更前被征求意见
• [ ] 有明确的停止条件（指标连续恶化/达到目标）