liveBook · Manning

Governance

As explained in Architecting Autonomous Agentic Systems

The system of oversight, policies, controls, and accountability used to direct and constrain organizational systems and decisions.

At 3 a.m., a fraud agent may need to contain a credential compromise before a human can respond. Governance determines what it may do, when it must escalate, and who remains accountable. It combines oversight, policies, controls, and accountability to direct and constrain decisions. section 1.5.1 section 4.6

Traditional automation can often be reviewed against a fixed path before deployment. Autonomous systems select actions at runtime, so upfront approval alone is insufficient. Governance must remain close enough to execution to constrain, redirect, or pause behavior—not merely observe it. section 1.2.1 section 2.1.4

Bounded fraud response

Define account-freezing authority up to a specified number of customers and require human sign-off above that threshold. Express the compromise-probability policy as a machine-readable signal, and give the invoked application an explicit contract for what the agent may request and what confirmation is expected. section 1.5.1

Make governance operational at each decision boundary:

  • Specify delegated authority and escalation triggers, including confidence thresholds, policy violations, and consequence thresholds. Name the business owner accountable for the decision. section 4.6 section 4.5.1
  • Apply policy constraints before execution and record the action, policy, and outcome for audit and review. Use scoped permissions and time-limited credentials to limit exposure. section 1.3
  • Maintain the ability to change thresholds, redirect flows, or pause autonomy while it is live. Observation without intervention is not effective control. section 2.1.4
  • Review reasoning traces, refine policies as conditions change, and stress-test scenarios before expanding authority. section 3.3.3

Governance is a recurring operating cost. Tighter authority boundaries can reduce a failure’s blast radius, but may increase escalations and manual review. A risk-adjusted business case should account for both the controls and the time needed to establish trust. section 3.3.3 section 3.5.1

Annual governance cost ($/year) = (Governance hours per week × Loaded hourly rate ($/hour) × 52 weeks) + Annual simulation compute cost ($/year) + Annual governance tooling cost ($/year)

Estimate annual governance cost from staff time, simulation compute, and tooling rather than treating governance as a one-time expense. section 3.3.3

Pilot teams often rely on shared judgment and informal escalation. At scale, make authority, ownership, detection, escalation, and intervention explicit before transferring more decisions to an agent. Gradual delegation and parallel runs can control that transfer. section 4.5.1 section 4.6

How Architecting Autonomous Agentic Systems covers it

How deep it goes
Develops operational controls, escalation ownership, ongoing review, and governance cost estimation, with practical decision-boundary examples.
Examples
Illustrates governance through a fraud agent’s account-freezing threshold and a financial-services data-handling error under narrow versus wide autonomy.
What you can do after
Apply governance by defining authority boundaries, escalation triggers, accountable owners, and runtime intervention mechanisms.
How it connects
Connects autonomy with policy-as-code, observability, delegated authority, escalation, and enterprise architecture.

Also explained in

Compare how 9 books explain governance →

Related concepts

This explainer was generated by Manning's knowledge graph from the text of Architecting Autonomous Agentic Systems. The section links point to the parts of the book it draws on.