
Designing Agents That Close Loops, Not Just Tasks
There’s a useful distinction between agents that complete tasks and agents that close loops. It’s worth designing for, as long as you’re honest about what closing a loop asks of the people around it.
The difference between tasks and loops
Tasks: individual actions. An agent might send an email, generate a report or book a meeting.
Loops: the whole process from trigger to result. A loop starts with a need, involves whatever tasks that need requires, and ends only when the intended result has actually happened, with evidence that it has.
The gap between the two is where balls get dropped. The email was sent, but nobody replied. The report was generated, but the decision it was for never got made.
Why closing loops matters, and what it doesn’t buy you
A system that follows work through to its result, rather than stopping at the last action it took, can catch the follow-up that would otherwise slip. That’s the appeal.
What it doesn’t buy you is freedom from oversight. A closed loop is a claim that the result happened. Someone has to be accountable for that claim being true, and the more the loop can do on its own, the more a wrong closure can cost.
Anthropic’s engineering post Demystifying evals for AI agents (January 2026) draws a distinction that helps here. It separates the transcript, the record of what an agent said and did, from the outcome, the actual state of the environment when it finished. Its example is a booking agent: the agent saying the flight is booked is part of the transcript, while whether a reservation exists in the system is the outcome. It also separates succeeding at least once across several attempts from succeeding on every attempt. A loop is closed only when the outcome is there, and a loop is worth trusting only when it closes correctly consistently.
An illustrative loop
Illustrative example, invented. A customer reports a billing error.
- Trigger: a support ticket tagged as a billing error.
- Work: the agent finds the invoice, checks it against the customer’s plan, and drafts a correction and a reply.
- Result: the customer’s account shows the corrected amount, and the customer has confirmed it or hasn’t disputed it within a week.
- Evidence: the corrected invoice in the billing system and the ticket thread. “I sent the customer an update” isn’t evidence of the result; the corrected invoice is.
- Accountable reviewer: the billing lead, a named person rather than “the team”, who owns whether closures are right.
The simpler alternative is assistance. The agent does the lookup and drafts the correction and the reply, and someone on the billing team approves both before anything changes. For plenty of teams, that’s the right place to stop.
Setting review by the cost of a wrong closure
Review doesn’t disappear when a loop can close itself. It changes shape, and the shape should follow what a wrong closure costs:
- Every closure when a mistake is expensive, reaches customers or is hard to fix. In the example, any refund or credit is approved by the billing lead before it’s applied, not checked afterwards.
- A sample when mistakes are cheap and easy to reverse, and the record shows closures are being done correctly. The billing lead might review a fixed share of closed tickets each week.
- Exceptions only when there’s a long record of correct closures, a wrong closure is cheap to undo, and the system reliably flags the cases it isn’t sure about.
Moving down that list is earned, not assumed. Reduce review only when there’s a record of correct closures, checked against outcomes rather than the agent’s own reports, and a wrong closure is cheap to reverse. Actions that are expensive or hard to reverse, such as issuing money, cancelling a contract or telling a customer something you can’t take back, need approval before they happen, however good the record is.
When a closure turns out to be wrong, the loop reopens: further actions on that case stop, the accountable reviewer is told, and the work is corrected.
When not to hand it over
For the billing loop, keep a person approving, or stay with assistance, when:
- the correction issues a refund or credit, or changes what the customer will be charged in future
- the customer is disputing the charge, upset, or talking about leaving
- the plan or contract terms are unusual enough that the agent has nothing reliable to check against
- there’s no record yet of the agent getting these right
- nobody has been named as the accountable reviewer
Designing agents that close loops
- Define done as a result with evidence. The agent needs to know the intended outcome and where the proof of it lives, not just the task.
- Check the outcome, not the transcript. Build in a way to confirm the result happened: a record changed, a reply received, a confirmation from the person who asked.
- Grant authority within limits, and name who owns it. Spell out what the agent may do alone, what needs approval and from whom, and who reviews closures.
- Plan for the wrong closure. Decide in advance how a closed loop gets reopened and who hears about it.
Before you let it close anything
Pick one loop you’d like an agent to close. Write down the evidence that would prove it’s done, who reviews closures, and whether that review covers every closure, a sample, or exceptions only. Then run it through the autonomy test before it can send anything to a customer or move any money.
Updated:
Want help turning ideas like these into reality? Let's talk.