The agent, the platform around it, and the integrations it acts through. Every agent gets a bounded job, a permission boundary, an audit trail, and a path that still completes when a model is slow or wrong.
Next: Classification — The decision layer underneath.