Management by Exception
As agents do more work, you cannot watch every run or approve every action. The company has to bring you only what needs your judgment.
Founder, Task Machine
An agent that can act on its own creates a new problem the moment it starts working. You have to decide how much of what it does you actually need to see. Every run touches something. Every action could be the one you needed to catch. Multiply that by a dozen agents working through the day and the question stops being rhetorical.
Two answers present themselves, and both fail. You can watch everything, reading transcripts and confirming each step, until supervising the work costs as much as doing it would have. Or you can approve nothing and let the agents run, until the first quiet mistake teaches you why that was reckless. The founder who watches everything has bought a second job. The founder who watches nothing has bought a liability.
An older answer
There is an older answer. In 1911, Frederick Winslow Taylor described a principle he called management by exception. A manager, he argued, should not receive the full record of everything that went right. He should receive condensed, summarized reports that call attention to the exceptions, the results that fall outside the expected standard, both good and bad, so that his attention lands where it is actually needed. Everything within tolerance stays out of view.
The idea kept proving itself in other forms. Toyota's production system gave every worker an andon cord. The line runs on its own, and a worker pulls the cord only when something is wrong, which stops the line and brings help to the exact place it is needed. The cyberneticist Stafford Beer built the same intuition into his models of viable organizations, where each level absorbs the routine variety beneath it and passes upward only what it cannot resolve on its own. In every version the default is silence. The signal is the exception.
An agent-native company needs this more than a factory ever did, because agents generate exceptions faster than people can and hide them more easily.
The control surface is an inbox
This is why the control surface of an agent-native company is not a dashboard. A dashboard shows you everything and asks you to find the part that matters. That is management by inspection, and it scales as badly as watching every run. The control surface is an inbox: one place that stays quiet while work proceeds and speaks up only when a decision genuinely belongs to you. An approval before something irreversible. A question the agent cannot answer from what it has. A check that failed. A budget that ran out. A request for authority the work does not yet hold.
An inbox is not a feed of notifications, which interrupt for everything, and it is not a dashboard, which interrupts for nothing and makes you go looking. It is the short list of things that will not move correctly without you.
Three postures toward the work
The difference between the answers is easiest to see side by side.
| Posture | What reaches you | What it costs | How it fails |
|---|---|---|---|
| Watch everything | Every run and action | Supervision grows with execution | You bought a second job and capacity never grew |
| Approve nothing | Nothing, until something breaks | Nothing, until it costs everything | Quiet mistakes accumulate unseen |
| Manage by exception | Only what needs a decision | The work of defining what counts | Wrong thresholds surface noise or hide problems |
Only the third posture lets execution grow without a matching growth in demands on you. It is also the only one that treats your attention as the scarce resource it is.
An exception is useful only if you can act on it without going to dig. A good one arrives with what you need to decide: what happened, why the system could not resolve it alone, what will happen if you do nothing, and whether you have seen it before. The best ones also ask whether your answer should become a standing rule, so the same decision does not return next week wearing a different face.
What it costs
Management by exception is only as good as the standards that decide what counts as an exception. Set the bar too low and everything qualifies, which is watching everything with extra steps. Set it too high and real problems slip past in the silence.
Choosing what deserves your attention is not a setting you configure once. It is a judgment you keep making as the work teaches you where it tends to go wrong. That ongoing tuning is the real management, and no amount of automation removes it. What automation can do is make the tuning cheap: raise a threshold, add a check, turn a repeated answer into a rule, and watch whether the right things still reach you.
Being interrupted only when it matters
This is the posture Task Machine is built around. Approvals, unanswered questions, failed verifier checks, budget limits, and requests to expand an agent's authority arrive in one inbox. Routine work proceeds and leaves a record you can open when you choose to, not one that demands you read it. As a piece of work proves it can resolve its own exceptions, fewer of them reach you, which is the entire point.
You can measure a system like this by how little it needs to show you. A company that shows you everything has not automated the work. It has only moved the watching from the work to you.
What the inbox protects is the one thing agents cannot manufacture more of. Continue with Agent-Native Companies Need Better Judgment.