Production agent stacks have three components: observability, eval, and action. Your observability stack captures every tool call. Your eval suite judges whether the final output was correct. But the agent that runs tomorrow starts from a blank slate. The eval signal dies in a dashboard. This is the missing RL layer: Reflect sits between your evals and your agent. It treats traces not as passive audit logs, but as a training signal.