Organizational learning and retrospectives as a model for software agent learning:Many companies collect agentic traces ...

Organizational learning and retrospectives as a model for software agent learning:Many companies collect agentic traces of software engineering, and aim to use these as training materials for next generation of the models. But how to do it? The naive, classical method is to use rejection sampling and only leave in successful engineering examples, removing the failed ones.But is these a better way?Yes there is.You really have to think about these things in an epistemological frame — where the new knowledge and informative signal comes from for the next generation of models.It comes from not blind trial and error, but from learning about what worked and what did not, with hindsight.We can take an agentic trace, and then do a retrospective on it with AI agents. They can determine what worked, what did not, and they can rewrite the trace removing the bad choices, removing bad exploration decisions, rewriting the trace with hindsight, to produce an agentic trace which is superior to the cap...