"Levels of Autonomy for AI Agents" is an academic essay by Kevin Feng, David W. McDonald, and Amy X. Zhang (all University of Washington), published by the Knight First Amendment Institute (Knight Columbia) / Knight Columbia on July 28, 2025 and republished in the May 2026 Knight Columbia AI-in-Democratic-Society symposium. It introduces a five-level, user-centered framework for classifying the autonomy of AI agents and proposes "autonomy certificates" as a governance mechanism for single- and multi-agent systems.
Summary of argument
The essay argues that autonomy can be a deliberate design decision made by agent developers rather than an inevitable consequence of increasing agent capability. Given a fixed set of capabilities and a fixed operational environment, the authors hold that developers can make intentional choices about the level of autonomy with which an agent operates.
The framework is user-centered: it defines an AI agent's autonomy as "the extent to which it is designed to act without user involvement" and builds five levels around the role a user (human or AI) takes when interacting with an agent.
Terminology
The essay defines its core terms as follows:
- AI agent: a compound software system, inclusive of one or more AI models, that operates within an environment and takes actions within it. (Adopts the Russell-Norvig and Kapoor et al. compound-AI-systems framing.)
- User: an entity, human or AI, who issues an initial request for the agent's services. An agent can simultaneously be a user and be used.
- Agency: the capacity to formulate an intention for an action and carry out that action.
- Autonomy: the extent to which an AI agent is designed to operate without user involvement.
The authors distinguish agency from autonomy, which they say prior literature often conflates. An agent can have high agency but low autonomy (many tools but seeks user feedback), or low agency but high autonomy (a single calculator API but runs continuously in the background without user supervision).
The five levels
| Level | User Role | Description |
|---|---|---|
| L1 | Operator | User in charge at all times; agent provides on-demand support; "copilot" metaphor (e.g., Microsoft Copilot) |
| L2 | Collaborator | User and agent jointly drive workflow; agent may suggest plans, but user retains co-decision authority |
| L3 | Consultant | Agent drives execution; user consulted on key decisions; agent escalates ambiguity |
| L4 | Approver | Agent drives execution and decisions; user approves at gates before commit; (e.g., Anthropic Plan Mode) |
| L5 | Observer | Agent operates autonomously; user observes outcomes, intervenes only when something goes wrong |
The authors argue that capability and autonomy are independent design dimensions. A capable agent with good benchmark scores can be deliberately deployed at L1 to elicit user feedback, while a less-capable agent can run at L5 on well-scoped simple tasks.
Autonomy certificates
The essay proposes that a third-party governing body issue autonomy certificates to agent developers, communicating the behavioral characteristics of an agent to other developers and agents. The authors describe three applications:
- Risk assessment — third-party visibility into autonomy level enables ex ante harm estimation.
- Safety framework design — Responsible Scaling Policy–style policies can condition deployment thresholds on autonomy levels.
- Multi-agent systems engineering — agent-to-agent interaction can be parameterized by the certified autonomy of each participant.
The authors describe this as structurally analogous to type signatures in multi-agent systems: an agent declares, via certificate, its autonomy class, and other agents or governance entities can reason about expected behavior.
Evaluation gap
The essay observes that current agent evaluation primarily relies on capability benchmarks, such as GitHub-issue resolution, browser-use, and computer-use. If autonomy is a design decision separable from capability, the authors argue, the two should be measured separately. They propose evaluations that involve user-agent collaboration on tasks, rather than purely autonomous task completion, to probe an agent's autonomous behaviors. The authors' Section 1 critique characterizes the prevailing benchmark culture as one that "narrowly focuses on autonomous task completion accuracy."
Relation to other frameworks
The essay cites the Responsible Scaling Policy framing of "low-level autonomous capabilities" as significantly higher risk, and positions its five-level scheme as a more granular categorization than the RSP's binary autonomy-yes / autonomy-no capability threshold. It cites Mitchell et al. (2023) for the argument that fully autonomous agents should not be developed, and frames the five-level scheme as a constructive alternative offering granular calibration rather than total prohibition. The L4 "Approver" level corresponds to Anthropic Plan Mode (see Agent Architecture Patterns).
Relationships
- supports: Levels of Autonomy for AI Agents (Feng-McDonald-Zhang framework)
- deploys-in: Agentic AI, Agent Architecture Patterns (Anthropic's architecture vocabulary), Agent Autonomy Spectrum (5 Levels)
- related: Building Effective AI Agents (Anthropic), Trustworthy Agents in Practice (Anthropic, April 2026) (Anthropic agent safety principles), Responsible Scaling Policy (RSP), Principal-Agent Problem Applied to AI, Autonomy Certificates
- part-of cluster: Knight Columbia 2026 AI-in-democratic-society symposium
Sources
Raw Sources/Levels of Autonomy for AI Agents.md- Published at Knight Columbia: knightcolumbia.org