Wizwand AutoExp (short for "auto experiments"), inspired by karpathy/autoresearch, is a one-line setup that turns any AI/ML training project into an autoresearch-style workflow, enabling AI coding agents (such as Codex or Claude Code) to run experiments automatically and optimize them iteratively.
In your project directory, send your coding agent (such as Codex or Claude Code) this message:
Read https://github.com/wizwand/autoexp/blob/main/make_autoexp.md and follow the instructions to create `autoexp_program.md` in the current directory.
What it does under the hood:
- Your coding agent will scan the project directory and infer the training command, evaluation metric, and other details from the codebase.
- It will then create a
autoexp_program.mdfile that defines how to run experiments automatically.
To kick off automated experiments, send your coding agent (such as Codex or Claude Code) this message:
Hi, have a look at autoexp_program.md and let's kick off the experiments!
What it does under the hood:
- Your coding agent will read
autoexp_program.mdand understand the training process and configuration. - It will run the baseline experiment and record the results in
autoexp_results.tsv. - It will then iteratively run experiments, keeping the best-performing ones and discarding the rest.
To customize the experiment setup, edit autoexp_program.md freely to adjust the configuration as needed.
- Autoresearch but autonomous experiment flow configured by AutoExp: wizwand/autoresearch-example/program.md
- Website: wizwand.com
- Follow on X: @wizwand_team
- Discord: Join our Discord
MIT