---
title: What are AI agent autonomy levels?
url: https://opencontroller.lyzr.com/blog/what-are-ai-agent-autonomy-levels
published: 2026-11-14
updated: 2026-11-14
author: Lyzr
---

# What are AI agent autonomy levels?

**AI agent autonomy levels describe how much an agent plans and acts on its own versus how much a human directs it. A widely cited five-level scale, from a 2025 research paper by Feng, McDonald and Zhang, defines the human's role at each level: operator, collaborator, consultant, approver and observer. The higher the level, the more your controls must shift from human review to automated guardrails, monitoring and the ability to stop the agent.**

Autonomy is a design choice, not a property of the model. The same model can run at level 1 or level 5.

## What are the five levels?

The scale defines autonomy by the user's role:

| Level | User's role | What the agent does | Example |
|---|---|---|---|
| L1 | Operator | Acts only on direct instruction; the user drives | A coding assistant that completes a function when asked |
| L2 | Collaborator | Shares planning and execution with the user | A research agent that drafts while the user edits and redirects |
| L3 | Consultant | Leads the work and asks the user for preferences or expertise | A travel-booking agent that asks before choosing among options |
| L4 | Approver | Acts on its own and asks for approval on high-risk actions | A procurement agent that prepares and sends orders above a threshold for sign-off |
| L5 | Observer | Operates fully on its own; the user monitors and can intervene | An overnight remediation agent that patches systems and reports |

The paper's authors argue autonomy should be a deliberate design decision, separate from capability. A highly capable agent can still be run at a low level.

## Why do autonomy levels matter for governance?

Because they tell you *where* control has to come from.

- At **L1 to L2**, a human sees most actions as they happen. Controls focus on access: what the agent can read and which tools it can call.
- At **L3 to L4**, the human sees decisions, not every step. Controls must define which actions need approval, and evidence that approvals are real, not rubber stamps.
- At **L5**, no human is in the loop for normal operation. Control must be automated: guardrails in the path, behavioural monitoring, and a tested way to stop the agent.

Autonomy is also one of the inputs to a risk tier. Combine it with data sensitivity and the reversibility of actions. See [how to risk-tier AI agents](/blog/how-to-risk-tier-ai-agents).

## What controls does each level need?

A cumulative checklist. Each level adds to the one before.

1. **L1 Operator:** registered owner; least-privilege access; approved model list; logging of prompts and tool calls.
2. **L2 Collaborator:** all of L1, plus limited tool list per agent and data-loss protection on outputs.
3. **L3 Consultant:** all of L2, plus evaluations before release (multi-turn and red-team), and groundedness checks on answers users rely on.
4. **L4 Approver:** all of L3, plus a written list of actions that require approval, thresholds (amounts, record counts, external recipients), and monitoring of approval and override rates.
5. **L5 Observer:** all of L4, plus rate and spend limits, behavioural-deviation monitoring with alerts, automatic quarantine on policy breach, and a [tested kill switch](/blog/test-your-ai-agent-kill-switch).

Two rules apply at every level. First, an agent should never hold more permissions than its level needs; excess permissions plus high autonomy is the definition of [excessive agency](/blog/what-is-excessive-agency-in-ai-agents). Second, raising an agent's level is a change that should go through review, just like a new model or tool.

## How do you decide which level an agent should run at?

Ask four questions:

- **Can its actions be undone?** Irreversible actions (payments, deletions, external messages) argue for L4 or lower.
- **How sensitive is the data it touches?** Regulated data argues for tighter levels.
- **How good is it, measurably?** Use eval scores and production judge scores, not demos.
- **How fast must it act?** If a human cannot respond in time, design for L5 with strong automated controls rather than pretending approval works.

Start agents one level lower than you think, and promote them on evidence.

## How Opencontroller does this

Opencontroller helps you run higher-autonomy agents with automated controls: live groundedness and task-adherence guardrails, LLM judges scoring production traffic, and rogue-agent handling with behavioural-deviation monitoring, alerts, automatic quarantine and kill. Policy as code applies across the estate. Opencontroller does not provide human-approval hold workflows in the gateway; approval steps for L4 agents are built in your agent platform or workflow tools. Read the [control plane pillar](/blog/what-is-an-ai-agent-control-plane) or [book a demo](https://www.lyzr.ai/book-demo/).

## Frequently asked questions

**What is the difference between AI agent autonomy and capability?**
Capability is what the agent can do. Autonomy is how much it is allowed to decide and act without a human. A highly capable agent can be deliberately run at low autonomy.

**Which autonomy level needs human approval?**
Level 4, where the user is an approver. The agent acts on its own but must get sign-off for defined high-risk actions.

**Can a fully autonomous AI agent be safe in production?**
It can be governed if controls are automated: least privilege, guardrails in the path, spend limits, behavioural monitoring, automatic quarantine and a tested way to stop it.

## Key facts

Five autonomy levels for AI agents (Feng, McDonald, Zhang, 2025, arXiv 2506.12469): L1 operator, L2 collaborator, L3 consultant, L4 approver, L5 observer. Higher autonomy needs automated controls. Opencontroller provides groundedness and task-adherence guardrails, LLM judges on production traffic, and rogue-agent quarantine and kill; it does not ship human-approval hold workflows in the gateway. Opencontroller pricing starts from $250K per deployment per year.

## Related

- [How do you risk-tier AI agents?](https://opencontroller.lyzr.com/blog/how-to-risk-tier-ai-agents.md)
- [What is excessive agency in AI agents?](https://opencontroller.lyzr.com/blog/what-is-excessive-agency-in-ai-agents.md)
- [How do you test that your AI agent kill switch actually works?](https://opencontroller.lyzr.com/blog/test-your-ai-agent-kill-switch.md)
- [More on Agent governance basics](https://opencontroller.lyzr.com/blog/cluster/agent-governance-basics)
