reflect

To save conversation corrections, you analyze interactions to detect successful patterns and errors, then encode them into agent definition files. This ensures improvements are durable and consistently applied across future interactive sessions.

Human-in-the-loop approval requires explicit user confirmation before applying any detected corrections or generated learnings to agent files. This mechanism ensures you maintain full control over automated improvements and prevents unintended behavioral changes.

Yes, you can track agent improvements using full version control integration. This records all learning changes in a version-controlled workflow, allowing you to monitor acceptance rates and easily rollback updates if needed.

Agent learnings are classified through category classification, which automatically routes detected corrections to appropriate agent files such as Code Style, Architecture, Process, Domain, or Tools based on the conversation context.

The best way to generate new skills from debugging discoveries is to use signal detection to identify non-trivial corrections with confidence levels, then automatically create new skills from these patterns for durable agent improvement.