Reinforcement Learnings

Turn reviewed production feedback into approved learning cases.

Reinforcement Learnings closes the loop between production behavior and future agent performance. It captures reviewed corrections from real usage and turns approved feedback into reusable learning cases.

What it does

Use Reinforcement Learnings to improve the agent from its own history. It helps teams collect issues found in production, review what the correct behavior should have been, and preserve approved corrections for future use.

This gives you a structured path from observed misses to durable improvements.

Approval flow

Learning cases do not enter memory automatically. Each case must be reviewed and approved before it is added.

Approval is manual. Auto-approval is intentionally disabled.

How it fits with the rest of Manage Agent

Use Monitoring and Online-Eval to find misses in production. Use Evals to validate improvements before you roll them out more broadly.

Reinforcement Learnings connects those steps. It carries reviewed feedback forward so the agent improves from real outcomes, not just isolated test runs.