autoresearch

# Autoresearch An autonomous agent that finds improvements through measured experiments or deep research. Based on Karpathy's autoresearch pattern: separate what the human controls (strategy) from what the agent controls (execution), then let the agent iterate indefinitely with objective verification. ## Choosing a Mode | Mode | Command | When to use | |------|---------|-------------| | **Optimize** | `/autoresearch optimize` | There is code/config/prompt + a way to measure quality. Find improvements autonomously. | | **Research** | `/autoresearch research` | Deep, multi-source research on a topic with synthesis. | | **Improve** | `/autoresearch improve` | Improve something without a clear starting point. Research best practices first, then apply via the optimize loop. | When no mode is specified, infer from context: metric or benchmark mentioned → Optimize. Question or topic exploration → Research. Wants something "better" without a defined measure → Improve. --- ## Mode 1: Optimize (Experiment Loop) The core Karpathy pattern. A hill-climbing ratchet where only measurable improvements accumulate. ### Step 1: Configure the Experiment Before looping, establish four components. Ask the user to confirm if anything is ambiguous — but if the project structure makes the answers obvious, just proceed. | Component | What it is | Example | |-----------|-----------|---------| | **Truth Layer** | Read-only files that define correctness — tests, specs, data, eval harness. The a...