autoresearch
Installation
SKILL.md
autoresearch
An autonomous optimization loop where subagents iteratively modify an artifact, evaluate it against a single scalar metric, and keep improvements. Inspired by Karpathy's autoresearch — generalized beyond ML training to any domain with a measurable objective.
The pattern in four invariants
| Component | What it is | Why it matters |
|---|---|---|
| Editable artifact | The file(s) the agent modifies each iteration | Bounds the search space |
| Evaluation oracle | A command that produces a single number on stdout | Unambiguous better/worse signal |
| Metric direction | minimize or maximize |
Tells the agent which way is "better" |
| Program document | Human-written intent, constraints, and boundaries | Steers the search without micromanaging |
The loop is simple: hypothesize → edit → evaluate → keep or revert → repeat.