When Execution Gets Cheap, Direction Gets Exposed 9 May 2026 7 min read
When agents do the work, fewer people stand between a leader’s brief and the result, and the repairs those people used to make stop happening.
Direction made visible Until recently, a leader’s direction passed through people before it became work. Those people asked what was meant. They noticed the constraint nobody had mentioned, guessed at the priority nobody had stated, and often fixed the brief without saying so. The finished deck or feature carried their repairs along with the original request, so from the outside a vague request and a clear one looked much the same.
An agent is far less likely to make those repairs. Given a gap in the brief, it usually fills the gap with a guess and returns work that looks complete. The gap surfaces later, as a system doing something nobody decided it should do, and the trail then leads straight back to the brief and to the person who judged it good enough to send. Cheap execution removes the cover that weak direction used to have, and the direction is judged on its own.
I read this as a correction. Organizations have long judged leaders by the work that came out at the far end, because the work was easy to see and the decisions behind it were hard to see. Now that an agent can produce the deck or the feature, finished work says less about the person who asked for it. Two questions take its place: who knew enough to act, and who owns what happened afterward?
The two obligations Those questions mark out the part of leadership this essay is about. A leader hands work to someone else: a strategy to a team, a request to an analyst, a brief to an agent. Before the handoff, the leader has to decide that the direction is clear enough for another party to carry out, and has to decide it while some facts are still missing. After the handoff, the leader has to answer for the result. I will call the first obligation the threshold, meaning the point at which you know enough to hand the work off. The second is ownership.
Leadership includes more than this. Leaders hire and develop people, hold a team together, and make the case for where the organization is going. I am setting those aside because the threshold and ownership are the obligations that change when the party receiving the work is an agent. With fewer people between the direction and the result, a threshold set too low is caught less often, and ownership is harder to spread across a team.
Enough to act You have met the threshold when you can say six things in plain words: what the goal is, what constraints apply, where the main risks are, what you are trading away, how the work is most likely to fail, and how you will recognize success or failure. If you cannot say them, whoever does the work will supply the missing answers for you.
Leaders get the threshold wrong in two directions. Some set it too high and wait for complete information, which rarely comes; while they wait, customers and competitors make the decision for them. Others set it too low. They act on a hunch, or on a sense that the room agrees, and export randomness to the people who have to absorb the consequences.
How high to set the threshold depends on what a mistake would cost and whether it can be undone. A change to a button label can go ahead on little evidence, because it can be reversed in a minute. A change to billing, or to who can see customer data, needs more, because a mistake there reaches people and is slow to undo. Work the organization has done many times needs less evidence than work it has never done. Two checks help before sending. Assume the work has failed and ask why. Then ask whether you believe you have enough because the evidence says so, or because you want to ship, avoid a conflict, or stop thinking about it.
What the agent has to guess Sending work to an agent adds a third check: where will the agent have to guess, and what does each guess cost if it is wrong?
Some guesses are cheap. They are local and reversible: a variable name, a log line, the order of two independent operations. A wrong cheap guess creates cleanup. Other guesses are expensive because they change the shape of the system: a schema choice, an authorization boundary, a billing rule, a swallowed exception. A wrong expensive guess produces a different system from the one the leader thought they had authorized.
Take a brief that says, “Let staff apply a discount at checkout.” It leaves open who counts as staff, whether a discount can exceed the price of the item, and whether the discount appears on the customer’s receipt. An agent will usually settle all three itself in order to finish, and the feature will run. If its answer to the first question is “any logged-in user,” the brief has produced an authorization rule that nobody chose.
Agents do ask clarifying questions. They ask fewer than a team would, and they rarely ask the one a senior colleague might: whether this should be built at all. So the work of finding the open guesses falls to the person writing the brief.
The amount of context a brief needs depends on the cost of the guesses it leaves open. A reversible design tweak may need one sentence. A database migration needs its constraints written down, a test that proves the change, and a named person ready to review the result. Telling the two kinds of task apart is the leader’s job. If the open guesses are expensive, the threshold has not been met and the brief should wait.
Ownership Ownership means the decision to send the brief stays yours once the results are in. If you sent it, you bought the outcome. You cannot say afterward that you always had concerns, and you cannot shelter behind an ambiguity you left in place. Some failures are bad luck, and owning those means only that you accept the decision to proceed as yours.
Handing work to people made ownership easy to avoid. A leader gave direction, a team interpreted it, the work changed along the way, and by the time the results came in, enough people had shaped the work that each could remember the original decision differently. A brief sent to an agent leaves less room for that. It sits closer to the leader’s judgment than a team’s interpretation ever did.
AI also offers a new way to avoid ownership. I call it accountability laundering, and it sounds like this: the model hallucinated, the agent went off the rails, the automation deployed the change. Each statement may be accurate, and none of them answers the questions that matter. Who set the agent loose? Who judged the brief clear enough? Who decided the guardrails would hold? Who approved the way the work was checked? The agent worked inside limits that a person set, and that person owns the decision to trust them.
Credit and blame have to follow the same rule. When an agent’s work succeeds, the leader earns credit the way a sales leader earns it for a strong quarter without closing every deal, by setting the conditions under which the work could succeed. When the work fails because the goal was vague or the check tested the wrong thing, the same person earns the blame. A leader who accepts the first and refuses the second is keeping two sets of books.
Effort used to soften this accounting. “I worked hard on this” was a defense when the work took weeks, and it carries little weight when an agent did the work in ten minutes. What a leader can still claim is judgment: I knew enough, I acted, I said what the work was, and I accept what followed. That standard is harder to fake than effort.
Continue the argument Practical workflow How I Build with AI Agents The operational counterpart: lanes, slices, and the review discipline that gives the threshold and ownership somewhere to live. Calibration essay When the Answer Sounds Finished Why fluent answers can close questions before evidence supports closure, and what calibrated abstention requires.