Autonomy is often described as a single switch: either a system acts or a person does. Real operating environments need a richer model. The same worker may be trusted to send an approved reminder, required to request review before changing a payment arrangement, and forbidden from making a regulated decision.
Govern the action, not the label
“Autonomous agent” is too broad to be useful. Governance should be attached to specific effects: reading a record, changing a field, contacting a customer, committing spend, offering terms, or publishing a decision.
For every tool action, choose a policy: allow, deny, allow with review, or require approval before execution.
Use risk as routing context
An approval request should arrive with the work objective, requested action, reason, risk type, supporting facts, relevant artifacts, and a clear expiry. A human should not have to reconstruct the entire journey to make a decision.
Design exception ownership
Workers will encounter unavailable resources, conflicting data, failed providers, ambiguous intent, policy limits, and external delays. Each exception needs an owner and a permitted recovery path. Otherwise “human in the loop” becomes an unstructured queue no one operates.
Preserve evidence by default
Every consequential action should leave enough evidence to answer four questions: What happened? Why did it happen? Which policy applied? Who or what authorized it?
This is essential for regulated work, but it is equally valuable for ordinary operations because evidence converts disputes and debugging into inspectable facts.
Promote behavior, do not silently mutate it
Reviewed outcomes may suggest better behavior. That change should become a candidate, be tested, compared, approved, versioned, and capable of rollback. The system should make improvement visible rather than allowing prompts to drift invisibly in production.
A practical maturity path
- Observe: the worker prepares work while a human executes.
- Approve: the worker acts after explicit review.
- Bounded autonomy: routine actions run automatically; defined risks escalate.
- Adaptive operation: evidence-backed improvements are tested within protected boundaries.
The goal is not maximum autonomy. It is the highest safe autonomy justified by the operating evidence.