Self-healing agents are agents that can detect certain failures and recover without waiting for a developer to intervene. Examples include retrying a failed tool call with corrected arguments, switching to a fallback retriever, asking for clarification, or rolling back to a known-good prompt.
Self-healing should be scoped to known, testable recovery paths. If the agent rewrites its own behavior without evals, baselines, and policy gates, it is not self-healing. It is unreviewed change.