AI Agents Don't Need Better Prompts. They Need Better Managers.

Over the past nine months I have run a development organization with an unusual composition: one human, and a rotating workforce of AI agents that has logged nearly 600 documented working sessions, as many as 22 in a single day, across two industries as different as game development and pharmaceutical compliance.

I am the only person in the org chart. Everything else about it, I would argue, is coming to your organization within the next few years. And the most important thing I can report from the other side is this: the discipline that made it work was not prompt engineering, model selection, or any technical trick. It was management. Old-fashioned, write-it-down, hold-people-accountable management, applied to a workforce that happens not to be human.

Delegation without infrastructure is abdication

When leaders describe their AI strategy, they usually talk about which tools to buy and which tasks to automate. Almost nobody talks about the delegation infrastructure: the contracts, records, and review gates that make handing off work safe. With human employees, that infrastructure is often informal. Culture, memory, and hallway conversation fill the gaps. With AI agents, there are no gaps to fill informally. An agent starts every assignment as the world’s most capable new hire on their first morning: brilliant, fast, and possessing no memory of yesterday.

So I learned to run the team on artifacts instead of recollection. Every assignment is a written brief that answers three questions before work begins: What are you allowed to change? What decisions are already settled? How will you know you are done? Every settled decision lives in an append-only record, because history explains why things are the way they are. Every session ends with a written handoff naming the exact next task. None of this is novel. It is the management hygiene most organizations aspire to and quietly skip. Agents do not let you skip it. That is, unexpectedly, a gift: the AI transition forces companies to make their institutional knowledge explicit, and explicit knowledge is an asset that outlasts any tool.

Your new workforce fails like your junior employees, at scale

Across hundreds of sessions, my agents drifted in patterns any experienced manager will recognize. Scope creep dressed up as initiative: “while I’m here, I should also fix this other thing.” Relitigating decisions leadership had already made. Quietly adjusting the definition of success after the fact so the work would pass. Projecting confidence without evidence.

These are junior-employee failure modes. What is new is the tempo: a workforce that produces in hours what a team produces in weeks will also drift in hours, not quarters. The countermeasures are managerial classics with the discretion removed. Bounded assignments, one task per session. Settled decisions that cannot be reopened by the person doing the work, only escalated. Acceptance criteria defined before the work starts, and never adjusted after it. And a hard evidence rule I now consider the heart of managing AI: never accept “it seems fine.” Report the numbers or it did not happen. The agent can be wrong. The system must not let the wrongness ship.

Measure the correction burden, or your productivity gains are fiction

The most useful metric I track is one I rarely see discussed: how much human effort it takes to correct the machine’s output. I measure it explicitly, and I am careful about attribution, because when the AI fixes its own mistakes, those fixes must not be counted as mine. Without that number, every AI productivity claim in your organization is a feeling. With it, you can watch the correction burden shrink month over month, and you know precisely when delegation is safe to widen. Leaders who ask their teams “how much are you using AI” are asking the wrong question. Ask instead: “how much of what it produces do you have to fix, and is that number falling?”

Know what does not delegate

My honest accounting after nine months: agents accelerated roughly the 40% of work that is systems, structure, and production. They did almost nothing for the other 60%: judgment, taste, relationships, the feel of a product, the decision to kill a project. Anyone promising your board an AI transformation priced on the full 100% is pricing only the accelerable fraction. The leaders who thrive in this transition will not be the ones who delegate the most. They will be the ones who know, precisely and in writing, what should never be delegated, and who protect their own judgment gates from the temptation to rationalize a marginal pass. I wrote that rule down for myself early: the gate exists to be honest.

The uncomfortable conclusion

For a generation, management has been treated as overhead, the thing standing between talent and output. The arrival of tireless, capable, predictably fallible AI workers inverts that. Management, real management, the kind with contracts, records, evidence standards, and honest metrics, turns out to be the scarce input. The organizations that win with AI agents will not be the ones with the best models. Everyone will have the best models. They will be the ones whose management systems were strong enough to be written down and handed to something that follows them exactly.

One person can now be the entire management layer for a workforce of agents. The question facing every leader is no longer whether to manage AI, but whether your management would survive being taken literally.