https://github.com/mac2bua/autoalgo
Autonomous algorithm optimization research - 19x speedup achieved through self-driven iterations
https://github.com/mac2bua/autoalgo
algorithm-optimization autoresearch numpy performance python
Last synced: 7 days ago
JSON representation
Autonomous algorithm optimization research - 19x speedup achieved through self-driven iterations
- Host: GitHub
- URL: https://github.com/mac2bua/autoalgo
- Owner: mac2bua
- Created: 2026-03-14T19:32:18.000Z (26 days ago)
- Default Branch: main
- Last Pushed: 2026-03-15T21:30:57.000Z (25 days ago)
- Last Synced: 2026-03-15T21:55:02.736Z (25 days ago)
- Topics: algorithm-optimization, autoresearch, numpy, performance, python
- Language: Python
- Homepage:
- Size: 23.4 KB
- Stars: 0
- Watchers: 0
- Forks: 0
- Open Issues: 0
-
Metadata Files:
- Readme: README.md
Awesome Lists containing this project
README
# autoalgo
> Autonomous algorithm optimization — an experiment to have the LLM do its own research on algorithm performance.
Inspired by [Andrej Karpathy's autoresearch](https://github.com/karpathy/autoresearch), this project applies the same autonomous research loop to algorithm performance optimization.
## Overview
This system autonomously optimizes algorithm implementations for runtime performance:
1. Runs a fixed set of algorithm benchmarks
2. Tries out different algorithmic approaches
3. Keeps improvements, discards regressions
4. Runs continuously until manually stopped
Each experiment takes ~5 minutes. Leave it running overnight and you can run ~12 experiments per hour.
## Running Unattended
To run the agent indefinitely without manual intervention, launch Claude Code with the dangerous mode flag:
```bash
# Ollama (author's setup)
ollama launch claude --model qwen3-coder-next:cloud -- --dangerously-skip-permissions
# Claude CLI
claude --dangerously-skip-permissions
```
**Note**: This grants the agent full permission to execute commands without approval. Only use in a dedicated project directory.
## Setup
1. **Create an experiment branch** (date format: DD-MM-YYYY):
```bash
git checkout -b autoalgo/15-03-2026
```
2. **Install dependencies**:
```bash
uv sync
```
3. **Run the baseline**:
```bash
uv run evaluate.py > run.log 2>&1
```
4. **Record the baseline** in `results.tsv`:
- Get the `total_time_ms` and `memory_mb` from `run.log`
- Add a row with status `baseline`
5. **Kick off the loop**:
Point Claude Code or another coding agent at `program.md` and let it run the loop.
## Experimentation
- **`algorithms.py`**: Modify this file with experimental optimizations
- **`evaluate.py`**: Fixed benchmark harness (DO NOT MODIFY)
- **`results.tsv`**: Tracks experiment results
Each experiment runs for ~5 minutes, testing different algorithm approaches.
## Files
| File | Purpose |
|------|---------|
| `algorithms.py` | Algorithm implementations (MODIFY THIS for optimization) |
| `evaluate.py` | Fixed evaluation harness (DO NOT MODIFY) |
| `results.tsv` | Results tracking |
| `program.md` | The autonomous experiment protocol |
## License
MIT