Choose autonomy per task. Define what the agent may decide and execute, and when it must escalate. Autonomy is a design choice separate from the agent's capability and operating environment, so you can change who decides without changing what the system can do.1 Treating autonomy as binary is one of the common mistakes teams make.2 Start with the work the agent will touch, set the approval point for each task, and raise autonomy only where the boundary holds.
Start with the task
Name the work before choosing the level. The right level depends on the task,3 and decision boundaries state what the agent may decide alone and what requires escalation.4
Given fixed capabilities and a fixed operating environment, developers can choose an autonomy level deliberately and design the agent around it.5 Write the boundary plainly: what the agent may prepare, send, or change, and which cases must reach you.
Use five autonomy levels
Use these levels as working settings for individual tasks. A single workflow can use different levels for research, message approval, follow-up, and reply handling.
The framework defines five escalating levels of agent autonomy.6 It names them operator, collaborator, consultant, approver, and observer.7
| Level | How to use it | Use it when |
|---|---|---|
| Operator7 | You give the instruction and initiate each action. | The task carries judgment, reputation, or unusual context. |
| Collaborator | The agent proposes the next move and you choose whether it happens. | You want help with decisions while keeping control of execution. |
| Consultant | The agent finds patterns, options, or recommendations and you take the action. | The agent can analyze the work, while the decision still needs your read. |
| Approver | The agent prepares the action and waits for your release. | The action is repeatable, though a person should inspect it before it leaves. |
| Observer | The agent executes autonomously within defined boundaries.8 You review exceptions instead of every routine action.9 | The task has clear inputs, safe actions, and an escalation path. |
Assign a level to each task
Map each level to an individual responsibility before turning on sending or reply handling. Start with the lowest level that removes work without hiding a decision you still need to make.
| Task | Starting level | What the agent can handle | Your boundary |
|---|---|---|---|
| Research and monitoring | Observer | Let the agent run research and monitoring.10 | Keep account choice and interpretation with you until the output is reliable. |
| Notes, preparation, and first drafts | Approver | AI should handle notes, preparation, first drafts, and messy follow-up work while judgment stays with you.11 | Review the substance and release the message yourself. |
| Outbound message decisions | Consultant | The workflow can include deciding what to write in outbound messages.12 | Assign ownership for research, approval, external actions, and live replies before the agent can act.13 |
| Follow-up timing and channel changes | Approver | The agent can decide when to follow up, when to change channel, and when to stop outreach.14 Follow-ups are increasingly sent automatically.15 | Approve the rules for timing, channel changes, stopping, and exceptions before moving to Observer. |
| Reply interpretation and qualification | Consultant or Approver | The agent can interpret a reply as a positive response, soft objection, out-of-office message, or hard refusal and route it accordingly.16 | Keep judgment on replies that need context, and make the escalation destination explicit. |
| Live conversations and strategic accounts | Operator or Collaborator | Keep the person responsible for relationship-building, live conversations, and strategic accounts.17 | Let the agent prepare context, then take over before the interaction needs judgment or trust. |
Let the handoff follow the consequence of an error. Research can usually run further ahead than an external message, and a draft can run further ahead than a live reply. This lets you set different levels without making the whole system wait for approval.
Raise autonomy with an escalation test
Raising a level widens a permission. Test the boundary before widening it on routine work, ambiguous work, and work that should stop.
Ask these questions for every task:
- What decision may the agent make without asking?
- What action may it execute outside the system?
- What signal means it should stop?
- Who receives the case when the signal is unclear?
- What record shows what it decided and why?
Human-in-the-loop operation does not mean a person reviews everything.9 Use approval where review can change the outcome, and exception handling for routine actions that stay inside the boundary.
Keep capability and autonomy separate during the review. A more capable agent does not automatically earn permission to send, change targeting, or answer a live reply. Task-completion accuracy alone is a narrow evaluation target for agents,18 so inspect the quality of the decisions, the escalation rate, and the cases that reached the wrong boundary.
What not to do
Keep these mistakes beside the permission rules when you review the workflow.
- Do not choose a single autonomy level for the whole agent when the appropriate level depends on the task.3
- Do not treat autonomy as a binary choice.2
- Do not let the agent own external actions before you have assigned ownership for approval and escalation.13
- Do not use autonomous task completion as the only test of whether the workflow is working.18
- Do not make full autonomy the default for every task. Immediate harms can include scams and private-information leaks, while gradual harms can include human deskilling, loss of critical thinking, and societal disempowerment.19