How to Run AI Research Loops With PRAXIST

Asking an AI to fix your code once and hoping it got it right is a coin flip. PRAXIST turns that single shot into a research loop — explore, test, measure, learn, repeat — until it reaches your goal or runs out of budget.

It is open on GitHub, and the catch sits with you: you need a project that runs, a number you can actually measure, and a way to check the results. More experiments do not automatically mean better outcomes.

What You Need

• A runnable project you already have working.
• Measurable success criteria for what "better" means.
• A way to verify the results yourself.

Tutorial

1. Open https://github.com/sapientinc/praxist to get PRAXIST. It is the system that coordinates the experiments, not another coding assistant.
2. Point it at an existing codebase. That codebase is what the agents will investigate and try to improve.
3. Pick one goal you can measure. The source suggests making an algorithm run faster, cutting memory consumption, improving a model's performance, or finding a better implementation for a specific task.
4. Pair it with a coding agent like Codex or Claude Code. PRAXIST works alongside them as a research orchestration layer, so the agent does the writing and PRAXIST runs the loop.
5. Start the loop and let the agents explore. Multiple agents investigate different approaches in parallel, each solution is tested against your objectives, and the results become evidence rather than AI-generated claims.
6. Keep iterating until the goal is reached or the budget runs out. Successful findings shape the next round of experiments, so each pass builds on what the last one proved.
7. Review the Fair Source License terms before you use it commercially. It is not a standard open-source license, so read it first.

Wrapping Up

You now have a way to turn a one-shot AI answer into a measured, repeatable experiment on your own code. The groundwork matters more than the tool, though — a project that runs, a clear target, and a way to verify results are what make the loop worth running. Start with one small goal you can score, and let the evidence decide.