The separation that matters
An agent runtime decides how one task proceeds. A control plane defines what every agent is allowed to do, provides the credentials and policy needed to do it, and records what happened. Keep execution fast and local; keep governance consistent and inspectable.
Centralize the decision and evidence. Keep agent execution loosely coupled from any single model, framework, or tool protocol.
Core layers
Start with an agent registry and workload identity. Add a policy decision point, just-in-time credential broker, tool and model registry, approval service, telemetry pipeline, budgets, evaluation gates, and lifecycle controls.
Failure boundaries
Assume tools time out, models change behavior, approvals expire, and a task retries. Use idempotency, explicit timeouts, bounded retries, circuit breakers, and a kill switch that can revoke credentials and stop new runs.
Generate a system-specific version
Use the interactive builder to translate these principles into layers and controls for your fleet size, framework, model set, tools, identity, and data requirements.