Overnight Experimenter

Give an AI agent a codebase, a metric, and a time budget — let it run experiments autonomously overnight. You wake up to a log of what it tried, what worked, and what didn't.

Inspired by Karpathy's autoresearch, but general-purpose. Works for performance optimization, prompt engineering, config tuning — anything with a measurable score.

~1000 lines of Python. Zero external dependencies.

Install

pip install overnight-experimenter

Requires Python 3.10+ and a coding agent CLI (Claude Code or Codex).

Quick Start

# Create an experiment
overnight init my-experiment

# Edit the files:
#   program.md   — what to optimize + constraints
#   evaluate.sh  — script that outputs a numeric score
#   workspace/   — the files the agent can modify

# Run it
overnight run my-experiment --budget 8h --agent claude --direction minimize

# Check progress
overnight status my-experiment

# Generate an HTML report
overnight report my-experiment

How It Works

experiment/
├── program.md          # Instructions for the agent
├── workspace/          # Files the agent modifies
├── evaluate.sh         # Returns a numeric score
├── experiments.jsonl   # Auto-generated log
└── best/               # Snapshot of best version

The loop:

1. Agent reads program.md + experiment history
2. Agent makes ONE modification to workspace/
3. evaluate.sh runs → numeric score
4. If improved → snapshot to best/, log success
5. If not → revert to best/, log failure
6. Repeat until time budget is exhausted

Example: Python Performance

The included example optimizes a naive prime-finding algorithm:

# Copy the example
cp -r examples/python-perf my-experiment

# Run for 30 minutes
overnight run my-experiment --budget 30m --agent claude --direction minimize

Starts with trial division (~0.11s), agent discovers sieve of Eratosthenes, bitwise tricks, etc.

CLI Reference

Run Options

License

MIT