The Agent Autonomy Spectrum is a five-level framework for classifying AI agents by how much they act independently of human approval. It was introduced by Nita Farahany in Class 26 of her introductory course (December 8, 2025). Farahany uses the levels as scaffolding for governance choices, in particular for deciding at which point human approval should be mandatory.
The five levels
The framework runs from a system with no agency to one that operates for days without supervision. Each level is illustrated with a paradigm task and, where applicable, an example system.
| Level | Name | Behavior | Example |
|---|---|---|---|
| 1 | Reactive (no agency) | Responds to a query with no tools, actions, or planning | Classic chatbot |
| 2 | Tool-Using | Uses tools only to gather information, then synthesizes a response | ChatGPT with web browsing |
| 3 | Planning | Breaks a task into steps and proposes a plan for the user's approval before acting | Most commercial agents today |
| 4 | Executing | Searches, selects, acts, and reports back without per-step approval | OpenAI Operator |
| 5 | Autonomous | Operates for hours or days without supervision, making many compounding decisions | Amazon's reported "frontier agents" |
Level 1 (Reactive) has no agency. A user asks "What's the capital of France?" and the system responds "Paris." It uses no tools, takes no actions, and does no planning. Farahany describes this as a classic chatbot, helpful but not an agent.
Level 2 (Tool-Using) uses tools, but only to gather information. A user asks "What's the weather in Tokyo?" and the system searches the web, retrieves data, and synthesizes a response. Farahany places ChatGPT with web browsing at this level.
Level 3 (Planning) proposes actions but asks before taking them. A user says "Plan a trip to Tokyo" and the system breaks the task into steps, researches options, and presents a detailed itinerary for the user's approval. Farahany describes most commercial agents as sitting at this level today.
Level 4 (Executing) searches, compares, selects, and acts, then reports back. A user says "Book me a trip to Tokyo under $2,000" and the system books it and replies, for example, "Done. Confirmation #12345." Farahany places OpenAI Operator at this level. The Fowler egg incident, in which Operator purchased eggs without user approval, occurred at this level (see below).
Level 5 (Autonomous) operates with minimal supervision over extended periods. Amazon is reported to be building "frontier agents" designed to work for "hours or even days" without supervision, making thousands of decisions that each compound on the last.
Governance application
Farahany uses the spectrum to argue that a single regulatory regime cannot sensibly cover both Level 2 (passive web search) and Level 5 (multi-day autonomous operation), and that governance responses should vary by level. The framework is presented as an instance of broader AI governance design.
The governance posture Farahany associates with each level:
- Levels 2–3: largely acceptable without intervention; transparency about AI use is treated as sufficient.
- Level 4: the point at which the principal-agent problem begins to bite. Farahany argues human approval should be mandatory for financial transactions, communications committing the user to obligations, and irreversible actions.
- Level 5: Farahany argues for special licensing, enhanced monitoring (with parallels to the monitorability tax), and strict access-control accounting.
The Fowler egg incident
The Fowler egg incident is the canonical Level-4 failure mode within the framework. On February 7, 2025, OpenAI's Operator bought $31.43 of eggs without user approval. OpenAI stated that "our safeguards did not work as intended."
Relation to other frameworks
The principal-agent problem in AI (Principal-Agent Problem Applied to AI) intersects directly with the spectrum: traditional agency law assumes Level-3-or-below "human secretaries" who act under reasonable expectations and meaningful supervision, and Levels 4–5 break those assumptions.
Under the "intelligence replaces hierarchy" thesis of Pereyra and Harvey (Intelligence Replaces Hierarchy), the agent Spectre operates at Level 5, triggered not by human prompts but by system monitoring of incidents, bugs, customer feedback, and Slack.
The Anthropic/Clearwater alignment risk update (Alignment Risk Update) becomes most relevant at Levels 4–5, on the argument that the cost of uncaught misbehavior compounds with autonomy.
The framework treats agentic AI (Agentic AI) as the underlying capability that the levels measure.
A regulator has since adopted the same organizing device. The UK FCA's July 2026 Mills Review (The Mills Review: AI and the future of retail financial services (FCA, July 2026)) structures its analysis of retail financial services around an "AI autonomy spectrum," projecting movement along it by 2030 and describing the corresponding change in the human role as running "from operators close to each decision, towards collaborators, approvers and, eventually, observers." The review states that this shift requires "a clearer account of what human oversight actually involves," locating the definitional gap at the levels the spectrum labels 4 and 5.
Relationships
- introduced-by: Inside My AI Law & Policy Class 26: When AI Stops Advising and Starts Acting (Farahany, December 2025)
- depends-on: Agentic AI
- related: The Mills Review: AI and the future of retail financial services (FCA, July 2026), Financial Services — AI Deployment, Principal-Agent Problem Applied to AI, Intelligence Replaces Hierarchy, Alignment Risk Update, Monitorability Tax
- instance-of: AI Governance (umbrella)