Design stage · Reference

Autonomy Patterns

The steps and the stakes place any AI task. The pattern shows where the human sits.

Steps predictable → ambiguous Can someone write the steps down today?
Stakes minutes → money, trust, compliance What does a wrong answer cost?
LowStakesHigh
Predictable · High stakes

Copilot

Suggests inside the person's own tool. The person acts.

Autonomy
Suggest
Human
Does the work and owns the result
e.g.
Write production code
Ambiguous · High stakes

Gated Agent

Gathers context, proposes, then waits for a yes.

Autonomy
Act-with-confirmation
Human
Approves, edits, or rejects each action
e.g.
Process high-value refunds
Predictable · Low stakes

AI Pipeline

A system event triggers it. Nobody watches each run.

Autonomy
Autonomous
Human
Spot-checks and handles the exceptions
e.g.
Route inbound support tickets
Ambiguous · Low stakes

Autonomous Agent

Plans, picks tools, and runs to completion in a sandbox.

Autonomy
Act-and-notify
Human
Reviews the result before it leaves the sandbox
e.g.
Draft a research brief
PredictableStepsAmbiguous
Stakes set the ceiling. A condition from a works council, a regulator, or a compliance owner wins over the grid. Settle the stakes with the person who owns the cost of a wrong answer.