AgentKeel Optimization (Embedded)

Any deployed fleet agent can embed AgentKeel — an exact + semantic response cache, prompt/token compression, and CCR (compute/cost reduction) offload. It is fail-open and opt-in via AGENTWATCH_OPTIMIZE=1. On a cache hit the agent serves the response locally and skips the provider entirely, eliminating both token cost and latency. Savings telemetry flows to the Observe Platform’s Optimize surface as a direct-vs-AgentKeel comparison. See §7.21.