I Built a Control Plane for My AI Agent — Because It Kept Making the Same Mistakes

The author built Evolve, a control plane for autonomous AI agents, to solve three key problems: lack of visibility into agent activity, repeated mistakes across sessions, and delayed error detection. The system uses self-reporting APIs (heartbeat, discovery, review) and a layered knowledge base where an LLM scores and persists lessons, injecting relevant knowledge into the agent's prompt at startup to prevent recurring errors.