# autor3search

> Point your coding agent at your repo and come back to faster code. The agent proposes optimizations, a frozen harness measures each one, and the verdict is KEEP or DISCARD — every experiment measured, nothing taken on faith.

autor3search is an open-source (MIT) autoresearch harness for code performance.
You point a coding agent at a repository; it proposes one optimization per
experiment, and a benchmark harness the agent cannot reach measures each one and
answers KEEP or DISCARD. Tests are frozen and restored before every run, the
metric lives in a compiled binary, noise is discarded and regressions rejected.
Accepted changes land as one commit each; every experiment, including the
failures, lands as one row in results.tsv.

To use it, copy the agent prompt for your language and paste it into your coding
agent from inside the repository you want to optimize. The prompts are verbatim
copies of the "## Start here" block of each repo's README.

## Agent prompts

- [Every prompt, verbatim, in one file](https://autor3search.dev/llms-full.txt): the copy-paste
  starting prompt for all 7 ports
- [go prompt](https://github.com/autor3search/go#start-here): the README block it is copied from
- [typescript prompt](https://github.com/autor3search/typescript#start-here): the README block it is copied from
- [javascript prompt](https://github.com/autor3search/javascript#start-here): the README block it is copied from
- [python prompt](https://github.com/autor3search/python#start-here): the README block it is copied from
- [java prompt](https://github.com/autor3search/java#start-here): the README block it is copied from
- [c# prompt](https://github.com/autor3search/csharp#start-here): the README block it is copied from
- [rust prompt](https://github.com/autor3search/rust#start-here): the README block it is copied from

## Ports

- [autor3search/go](https://github.com/autor3search/go): go install …@latest — early but working
- [autor3search/typescript](https://github.com/autor3search/typescript): npm install --save-dev — on npm · measures your *.bench.ts benchmarks
- [autor3search/javascript](https://github.com/autor3search/javascript): npm install -g — on npm · standalone, measures vitest benchmarks
- [autor3search/python](https://github.com/autor3search/python): uv tool install · pipx — on pypi · early but working — validated on humanize
- [autor3search/java](https://github.com/autor3search/java): release jar + launcher — early but working — validated on org.json
- [autor3search/csharp](https://github.com/autor3search/csharp): dotnet tool install -g — on nuget · early but working
- [autor3search/rust](https://github.com/autor3search/rust): cargo install --locked — on crates.io · measures with criterion

## Docs

- [Case study](https://github.com/autor3search/go/blob/main/docs/case-study.md): an overnight run, p<0.0001, validated on go-humanize, mapstructure and google/uuid
- [Landing page](https://autor3search.dev/): what it does, a replayed run, the loop
- [karpathy/autoresearch](https://github.com/karpathy/autoresearch): the project this one takes after