Autonomy Is Not an End in Itself
An AI agent is meant to take work off people’s hands. That does not mean it should make every decision itself. Real business processes contain actions with different levels of risk: compiling information is not the same as sending an invoice, deleting a record, or publishing a campaign.
A good agent harness treats these differences explicitly. Low-risk, reversible steps can run automatically. Irreversible, costly, or legally significant actions stop at a defined approval point. This does not turn the person into a permanent supervisor. It makes them the decision-maker at the few points where their responsibility genuinely matters.
Four Criteria for Meaningful Approval
Approvals should not be added everywhere on instinct. Otherwise, the result is a sequence of clicks that may look modern but saves very little work. A clear risk model is more useful.
- Reversibility: Can the action be undone completely and quickly?
- Potential harm: What financial, legal, or reputational consequences would an error have?
- Uncertainty: Does the agent have sufficiently reliable data and rules?
- External impact: Does the result leave the company or change a system of record?
The Agent Must Prepare the Decision
An approval is useful only if the person does not have to repeat all the work. The agent should therefore provide a proposal that can be checked: What is supposed to happen? On what basis? Which data was used? What uncertainty remains? And what will change after approval?
Instead of “May I proceed?”, the message should say something like: “I created proposal A based on these five inputs. The price and recipient have been checked. Sending it to the customer is the only irreversible step. Approve now?” This turns an interruption into a quick, informed decision.
Approval Also Means Resuming Cleanly
After approval, the harness must resume the process at exactly that point rather than starting over or executing the action twice. This requires state, idempotent operations, and a traceable transition between “prepared”, “approved”, “executed”, and “verified”.
The same applies to rejections. A productive agent records not only “No”, but also the reason and the next permitted step. This does not mean that the model automatically learns. Instead, the process becomes more precise for the next run and easier for the team to understand.
The Best Automation Makes Responsibility Clearer
Human-in-the-loop is not a sign that the agent is too weak. It is an architectural feature. A good system fully automates the reproducible part and makes responsibility explicit wherever a decision has real consequences. This increases speed without moving control into an invisible black box.