Agents that learn
Correct an agent once, and the next one knows
Each project has a wiki that people and agents both write. Agents read it before they start. A retro you can run every week looks at where people had to step in, and writes each lesson down.
- run
- scorecard
- retro
- wiki
- next run
-
wiki
A wiki every agent reads first
How to run the tests, which page is fragile, what the team decided last month. Every agent starts with the wiki’s home page and searches it when it hits something new. Every edit is kept, and any one can be restored.
-
scorecard
A scorecard for every agent run
Each run records what it cost, how many tries its checks took, whether it opened the wiki, and the first thing a person said after it finished. That first message is often a correction.
-
retro
A retro on a schedule
Schedule a thread to run a retro every week. It reads the scorecard, writes lessons into the wiki, files issues for real bugs, and proposes changes to your project’s agent instructions as a pull request your engineers review.
-
your data
Lessons you can read
PatchWell doesn’t train a model on your code. What the agents learn is written in your wiki, where anyone on the team can read it, correct it or delete it.