liveBook · Manning

Control plane

As explained in Designing AI Agents

A control plane is the part of a system that governs its operation by managing authoritative decisions, policies, or configuration.

A control plane governs an agent system’s operation through authoritative policies, decisions, and execution controls. It is separate from memory: memory can preserve history and references, but stored information alone does not determine what may govern the next action or who may change it. section 4.0 section 9.6

A practical organizational control plane includes:

  • A registry of agents, owners, tools, versions, and lifecycle states. section 9.6
  • A policy service and a shared observation and audit layer. section 9.6
  • An execution layer with short-lived credentials, isolation, limits, and an independent kill switch. section 9.6

A prompt or hidden tool can steer an agent’s choices, but runtime enforcement is the security boundary. Governance also needs named responsibility: someone defines policy, someone operates the enforcement point, and someone answers for the outcome. Approval workflows must establish reviewer authority and retain decision evidence. section 9.6

Current decisions without lost history

A game team keeps decisions in an append-only ledger. New decisions supersede old ones instead of rewriting or deleting them. Memory retains the history and rationale, while the control plane exposes the current accepted decision to execution. This prevents retrieval from treating an obsolete decision as current guidance. section 4.0 section 4.1.2

Enforcing a spending cap

A production gate receives a host-owned snapshot with each proposal. It atomically checks and reserves the amount before execution, preventing concurrent proposals from spending the same remaining budget, then settles the reservation from the external result. An observer links the decision and snapshot to the effect. section 9.6

Governance costs latency, attention, and compute. Match the controls to the worst credible effect: reading a file needs little ceremony, while a production deployment may need multiple controls and a human checkpoint. The boundary should contain the action without consuming the autonomy it is meant to enable. section 9.6

Table 9.9 The four governance patterns. Each addresses a different part of the action lifecycle.
Pattern
Question
Topology
When
Approval Gate
Should this proposal happen?
Route
At proposal, resume, and pre-commit revalidation
Progressive Commitment
What evidence earns or revokes autonomy?
Chain
Before commitment and across runs
Blast Radius Control
What is the largest permitted effect?
Hierarchy
At assembly time and throughout execution
Observability Harness
What happened, and can it be attributed?
Orchestrate
Throughout the lifecycle

Compare the four governance patterns by the question each answers and where it acts in the action lifecycle. section 9.6

How Designing AI Agents covers it

How deep it goes
Describes control-plane components and accountability, and illustrates runtime enforcement with an atomic budget gate; tool-level listings do not implement that gate.
Examples
Uses an append-only decision ledger and a spending-cap gate with atomic reservations to illustrate authority and enforcement.
What you can do after
Evaluate control-plane boundaries and select governance controls proportional to an action’s credible effects.
How it connects
Connects agent memory and decision supersession with approval gates, progressive commitment, blast radius control, and observability.

Also explained in

Compare how 10 books explain control plane →

Related concepts

This explainer was generated by Manning's knowledge graph from the text of Designing AI Agents. The section links point to the parts of the book it draws on.