liveBook · Manning

Progressive trust

As explained in Designing AI Agents

A governance approach in which an agent receives increasing autonomy as it demonstrates reliable behavior.

Progressive trust treats autonomy as an operational state earned through observed behavior and reversible after incidents. Grant authority per action, scaling review to reversibility, scope, and potential impact. section 2.5

Enforce trust boundaries in the runtime through:

  • Approval gates triggered by the current trust level. section 2.5
  • Action classification, human override, and audit trails. section 2.5

Vary authority by action

An agent might read an approved repository without interruption, require confirmation before a shell command, and have no authority to deploy. section 2.5

Onboard → Evaluate → Promote → Monitor → Demote (if needed).

Use this lifecycle to promote capabilities after observed performance and demote them after incidents. section 2.5

How Designing AI Agents covers it

How deep it goes
Defines an observed-behavior lifecycle and runtime safeguards, but gives no quantitative thresholds for promotion.
Examples
Contrasts repository reading, shell commands, email sending, and production deployment to illustrate differing authority levels.
What you can do after
Operate/configure runtime permissions by granting low-risk actions automatically, requiring approval for higher-risk actions, and revising authority after incidents.
How it connects
Links trust to approval gates, action classification, human override, audit trails, and blast-radius control.

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.