liveBook · Manning

Autonomy-level classification

As explained in Designing AI Agents

The degree to which an agent can make decisions and take actions without human intervention or authorization.

Autonomy-level classification describes how much an agent may decide and execute without human intervention. It is a degree of delegated authority, not an on/off property: the relevant questions are what the agent may do, when a person must intervene, and what evidence would justify more authority. section 9.1.1 section 9.1

Table 9.1 The autonomy spectrum. Five operational levels and a research horizon.
Level
Name
Agent role
Human role
Example
0
Observe
Suggests only
Decides and executes
Shadow-mode review
1
Assist
Reads, analyzes, and proposes
Reviews and acts
Read-only review assistant
2
Supervised
Executes routine actions
Approves higher-risk actions
Coding agent with approval
3
Autonomous
Handles bounded workflows
Reviews outcomes
Bounded background agent
4
Delegated
Operates within an explicit charter
Monitors and audits
Production agent within charter
5
Self-governing
Sets its own objectives
Audits periodically
Research horizon; not deployed

This working scale distinguishes five operating arrangements and a research horizon. Levels 1–3 are the practical deployment range; level 5 is not a deployment target. The scale is not an industry-standard taxonomy. section 9.1.1

A level is not blanket permission. Every agent has a hard ceiling on resources, amounts, and actions. Within that ceiling, its authority can expand or contract with observed performance and the risk of the task. A reliable agent may need tighter limits for a new, irreversible action than for a familiar, reversible one. section 9.1.1

Assign authority at the action level using reversibility, scope, and evidence. Reading an approved repository might proceed automatically, a shell command might require confirmation, and deployment might be prohibited. Enforce these boundaries in the runtime, not just in the prompt. Too much authority risks damage before review; too little makes the human a confirmation bottleneck. section 2.5

Governance controls make changes in autonomy enforceable:

  • Approval gates route concrete proposals to allow, deny, or ask; blast radius controls limit the effects of an action even when it is authorized. section 9.9
  • Progressive commitment changes working authority as evidence accumulates, while preserving the hard ceiling. Start new capabilities under review, promote them after observed performance, and demote them after incidents. section 9.9 section 2.5
  • Maintain human override and an audit trail of actions, authorizations, and rejected alternatives. Observability connects decisions and execution to outcomes, providing evidence for trust changes and responsibility. section 2.5 section 9.9

Review both individual proposals and their cumulative direction. A sequence of individually acceptable actions can drift away from the delegated objective. Credentials and role-based access can establish who may call a service, but not whether a particular proposed action is appropriate now. section 9.1

How Designing AI Agents covers it

How deep it goes
Defines levels 0–4 as operating arrangements and level 5 as a research horizon, while distinguishing the working scale from an industry standard.
Examples
Illustrates levels with shadow-mode review, a read-only assistant, a coding agent requiring approval, and a bounded background agent.
What you can do after
Evaluate and understand an agent’s authority by naming its level, hard ceiling, task risk, and evidence for changes in trust.
How it connects
Connects autonomy levels to approval gates, blast radius control, progressive commitment, and observability.

Also explained in

Compare how 2 books explain autonomy-level classification →

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.