The control plane is the product
When AI agents chain actions — querying data, calling APIs, modifying records, triggering workflows — the control plane is not infrastructure. It is the product.
What a control plane provides
- Execution boundaries: What actions can this agent take? What is the maximum scope?
- Approval gates: Which actions require human approval before execution?
- Kill switches: How do you stop an agent mid-execution? How do you roll back?
- Rate limiting: How many actions per minute? How much spend?
- Replay and audit: Can you reconstruct exactly what happened?
The circuit breaker pattern
When an agent's error rate exceeds a threshold, the circuit breaker opens and stops further actions. A human must investigate and reset. This prevents a malfunctioning agent from causing escalating damage.
An agent without a control plane is a script with ambition. An agent with a control plane is a system that can operate in production.