Use normal work as feedback
One way to improve an agent is to collect its traces along with a signal of whether its actions were correct. In the example I was responding to, that signal came from whether a code change was accepted during normal PR review.
Feed those results, the traces, and the agent's context into a different model to look for improvements. Use what it finds to change the agent's context, then track correctness over time to see whether the changes help.
The useful part is that people can review PRs as they already do. Getting that feedback without asking them to take on another task is a design problem worth spending time on.
Originally posted on X. Lightly edited for reading here.