Control without becoming the bottleneck

The moment agents start doing real work, you face a tension: every approval you require keeps you in control, and every approval you require costs you attention. This page is about resolving that tension — keeping the decisions that need you while letting everything else proceed — because getting it wrong in either direction quietly kills an agent operation.

Both failure modes look responsible at first

The first failure mode is approving everything. It feels safe: nothing happens without your sign-off, so nothing can go wrong unseen. But if every draft, every step, every routine action waits on you, the agents' capacity is capped at your reading speed. You have not gained a workforce. You have gained a queue. Work stalls whenever you step away, and you spend your day rubber-stamping the same reliable output you approved yesterday.

The second failure mode is the blank check. Frustrated by the queue, you turn everything loose — and now work happens that you never see until it has already landed. You lose the thread on what your own company is sending, publishing, and spending. When something does go wrong, you find out from the outside, and you cannot say when the last good state was. Full trust granted all at once is not delegation. It is abdication.

Both modes fail the same way: they treat control as a single switch — all approvals or none — when the useful question is which decisions need you.

The inbox concentrates the decisions that need judgment

The middle path starts with the inbox. It is not a feed of everything agents did — it is the queue of things only you can resolve: an approval before something irreversible, a question an agent cannot answer alone, a finished draft that needs your read, a proposal awaiting your yes. Routine progress does not arrive there, because routine progress does not need you. When the inbox is calibrated well, working through it is running the company: every item is a genuine decision, and clearing it means every pending judgment has been made.

Approval, question, and verifier steps place you inside the work

The calibration itself happens inside workflows. A workflow is the defined shape of a recurring job, and three kinds of steps put a human exactly where judgment matters. An approval step halts the run until you say go — you place it right before the consequential action, like sending outreach, not in front of every step. A question step lets the run ask you something it genuinely cannot decide — which segment to target this week — and continue with your answer. A verifier step checks the work against criteria you set before it ever reaches you, so what lands in your inbox has already cleared a bar.

Everything between those steps proceeds without you. That is the design: research, drafting, formatting, and assembly run to completion, and your attention is spent only at the points you chose. Add the autonomy level on each agent and the budgets that cap spend, and control becomes a set of placed boundaries rather than a mood.

The three step types divide the checking work

The steps are not three flavors of the same interruption — they split the checking into the part a machine can do and the part only you can. The verifier carries the mechanical share: links resolve, sections are present, the draft holds to the voice guide — checks that would otherwise waste your reading time on work that was going to fail anyway. The question step moves ambiguity to where it is cheap: the run asks before drafting in the wrong direction, instead of you discovering the wrong direction in a finished piece. What remains for the approval step is pure judgment — whether this is the right thing to send, publish, or spend on — which is exactly the work you cannot delegate and should not dilute with proofreading.

Your inbox load is the calibration signal

You do not have to guess whether the boundaries sit in the right places — your inbox tells you. If you are clearing a queue of approvals where you change nothing, the checking is placed too early: raise that agent's autonomy level, or move the approval step later so it guards only the consequential action, and let the verifier carry what you were rubber-stamping. If, instead, work surprises you — an output lands somewhere before you expected to see it, or you learn about a decision after the fact — the boundaries are too loose: add an approval step where the surprise got through, or step the autonomy back down for that kind of work.

Read the load over a week rather than a day, since any single morning can be noisy. A calibrated inbox has a recognizable feel: every item makes you actually think, and nothing that mattered got past you. When you have that, leave the dials alone.

The three surfaces carry the operating rhythm

Day to day, the middle path is the surface triad in motion: you decide direction in Discuss, the decisions become tasks agents pick up, judgment calls return to the Inbox, and when a specific piece of work needs steering you step into its task with the full history in front of you. Discuss to decide, Inbox to approve, Tasks to steer — the loop keeps you in charge of outcomes without making you a gate on activity.

The natural next question is how quickly to loosen the boundaries you placed. That is what autonomy levels are for — Autonomy levels in practice is next.