Self-healing agents sound inevitable: an agent fails, understands what went wrong, proposes a fix, validates it, and improves the system. But without observability, self-healing is just guessing.
This talk explains why reliable agent improvement needs a feedback loop built on traces, evaluations, approvals, and rollback. We will follow an agent run across prompts, model calls, tool selection, MCP servers, validation, policy checks, downstream APIs, and final responses, then show what evidence is needed to reconstruct what happened.
The core idea is simple: agents can only improve safely when they can observe their own behavior. We will discuss what kinds of failures can be detected from telemetry, what changes an agent might propose, where evaluations fit, and which actions should require human or policy approval. The goal is to make “self-healing” concrete, bounded, and useful for production agent systems.
