Improving Agent Behavior Using Reinforcement Learning from Traces in Python

Most AI agents kind of work. They do the job on a good day, then make a baffling decision on a bad one, and you are back to rewriting the prompt for the hundredth time. Prompting alone gives you a ceiling on how much you can control behavior.
This talk shows a different lever: let the agent learn from its own runs. Starting from a working Python agent, we capture its execution traces, define what "good behavior" means using evaluation graders, and use reinforcement learning to reinforce the better decisions. The whole loop runs on open-source Python tooling and open-weight models, so it is reproducible and cheap enough to run outside a big lab.
We demo it end to end on Afara Dada Code, a real agent that connects donors to women-in-tech initiatives across Africa, where reliable behavior genuinely matters. You will leave with a concrete, framework-agnostic workflow for improving an agent you already have, without redesigning it from scratch.