Insight · August 20, 2026
The pause is
the product.
An agent that acts without asking is a liability. An agent that asks about everything is useless. Human in the loop design is where that line gets drawn.
01 · The shift
The interface stopped being the app. It became the pause.
A year ago, most AI features lived inside a screen a person drove. Type a prompt, read the output, decide what to do next. The interface was the app. The person was the loop.
Agents changed that. The work happens on its own now. The screen a person opens is not where the work is being done. It is where the work stops to ask. That screen is the product.
The EU AI Act is now the floor, not the ceiling. Article 14 requires high risk systems to be designed so a person assigned to oversight can understand the system, monitor for anomalies, correctly interpret the output, and decide not to use the output or override it. That is a product specification, and it applies to more than the systems the regulation names.
02 · The problem
Full autonomy is a demo. Full oversight is a queue nobody works.
The naive move is to route every agent action past a person. Everything is safe, and nothing ships. Reviewers turn into a rubber stamp inside a week. Automation bias takes over. In one study of AI assisted diagnosis, pathologists reversed a correct call to match a wrong AI suggestion in 7 percent of observations. When the machine is loud enough and the queue is long enough, people click approve.
The other move is worse. Let the agent act on everything, trust its confidence, and read the log later. Large language models are systematically miscalibrated. The training that makes them helpful also teaches them to sound confident when they are wrong. A confidence number is a starting point, not a green light.
Somewhere between the two is the product. A small set of moves that pause, a compact view that supports a decision in seconds, and everything else running on its own.
The one line to keep
“The interface for oversight is the product now, not a screen you bolt on later.”
03 · What it actually is
Five beats, no matter how large the system.
01
Classify the action
Every action the agent can take gets a class · read only, reversible, external, or high risk. The class decides whether it gates.
02
Interrupt before the risky move
The agent saves state, names what it wants to do, and pauses. Nothing external has happened yet. The pause is a first class step, not a timeout.
03
Compress the state into evidence
One small view · what was asked, what the agent found, what it plans, and how confident it is. Enough to decide in seconds, not minutes.
04
Offer the choice, not the log
Approve, edit, or reject. One tap for the common case, an edit for the exception. Reviewers cannot audit a paragraph, only a decision.
05
Log the decision so the next one is faster
Every approval and edit is a labeled example. The next threshold moves, the next evidence view improves, and the queue gets shorter on its own.
04 · Not a review queue
A queue collects. A pause interrupts.
A queue is a list of finished work waiting for someone to say yes. By the time the reviewer opens it, the moment to intervene is gone. Half the queue is stale, the other half is guessed at.
A pause is different. The agent stops before the action. It writes down what it was going to do, saves its state, and waits. In modern agent frameworks the interrupt is a first class step, not an exception. LangGraph made this explicit with an interrupt primitive and durable checkpoints keyed by a thread id, so the agent can pause for minutes or hours and resume exactly where it left off. The reviewer is not clearing a backlog. They are unblocking a specific move on a specific run.
The difference shows up on the calendar. A queue grows. Pauses do not. If more work is pausing than a person can clear in the moment, the threshold is wrong, the classes are wrong, or the agent should not be running at all.
05 · One example that shows the whole idea
A refund agent that pauses at the money.
A support agent handles refund requests. It reads the ticket, pulls the order, checks the policy, and drafts a response. Reading and drafting are read only. They run freely. Sending the reply is external. It goes ahead if the policy match is clean and confidence is high, and the agent logs it. Issuing the actual refund is where the money moves.
Any refund over the threshold pauses. So does any refund where the policy match is soft or the customer history is unusual. The reviewer sees one card. The request in plain language. The order, the policy, the two lines from the terms the agent is relying on, its recommended refund amount, and a confidence score. Approve, edit the amount, or reject with a reason. The whole thing takes about ten seconds. Every approval, edit, and rejection becomes a labeled example the system learns from.
action: issue_refund class: high_risk amount: 214.00 policy_match: partial · 60 day return, item shipped day 63 customer: 3 prior orders, 0 prior refunds confidence: 0.71 recommend: approve · policy allows 90 day window for damaged goods [ approve ] [ edit amount ] [ reject ]
An evidence card · designed for a decision, not a paragraph
06 · Four classes, one rule for each
The class of the action decides, not the confidence score.
Read only actions run freely. Fetching a record, reading a file, running a query that changes nothing. If the agent pauses here, the whole system stalls for no reason.
Reversible actions run, and a person can undo. Drafts, labels, notes, staged writes. The system tells the reviewer what happened after the fact, and the undo is one click.
External actions run after a person confirms once, with context on screen. Sending a message, calling a partner API, publishing to a channel. The pause is short because the evidence is already assembled.
High risk and low confidence actions always pause. Moving money, deleting data, granting access, anything that touches a regulated decision. Colorado's revised AI law, signed by Governor Polis on May 14, 2026 and taking effect January 1, 2027, requires meaningful human review and reconsideration for consequential decisions made by automated systems, with a designated person who has the authority to override. Confidence scores are useful for the middle of that list, not for deciding whether to gate at all.
07 · When not to build one
A pause without a decision is theater.
Some work does not need a pause. If the action is fully reversible, cheap to redo, and audited after the fact, a gate only slows the system down. A gate that never causes a reviewer to change the outcome is a rubber stamp with extra steps.
Some work should not be automated at all. If you cannot describe the finish line, if the reviewer has no way to tell whether the agent is right, the answer is not a better evidence card. The answer is a person doing the work.
And a pause without an owner is worse than no pause. Agents that stop for a review nobody is assigned to run fail silently in the middle of the night. Every gate needs a named human, a response window, and a fallback if the window is missed. Otherwise the gate is a way to make risk someone else's problem.
Closing
An agent you cannot pause is not yours.
Pick one agent that is already running. List every action it can take, sort them into the four classes, and design one evidence card for the class that gates. The first day a reviewer clears the queue in a few minutes and the money still moves, you will know the pause is not a cost. It is the product.
EU AI Act, Article 14 on human oversight · Colorado Automated Decision Making Technology Act, signed May 14, 2026, effective January 1, 2027 · automation bias data from published clinical AI studies · interrupt and checkpoint patterns from LangGraph documentation
Share this perspective
More insights
Adjacent perspectives.
Bttr. Field Brief
The brief Bttr. writes for senior buyers.
Monthly. One signal worth your time on Brand Operating Systems, AI search visibility, and the infrastructure buildout. No filler.