The Agent's Manager

Companies are hiring a new kind of workforce at a speed no human workforce ever grew: agents that draft, book, triage, negotiate, and execute. And they are onboarding this workforce with a diligence no human hire would ever receive: no manager, no probation, no performance review, no one whose job is to notice that a worker has gone quietly wrong. If a human employee operated the way most deployed agents operate, unsupervised, unreviewed, answerable to a Slack channel at best, you would call it a governance failure. We call it automation.

The org chart with a hole in it

Draw your actual org chart, including the agents, and be honest about the reporting lines. The customer-service agents report to, in practice, a vendor dashboard. The procurement agent reports to whoever last edited its prompt. The research agents report to nobody, because they mostly work, and mostly working is how systems earn the right to fail unobserved. Every human box on the chart has a line going up. The agent boxes float. I wrote about the vacant seat in Who Answers When an AI Agent Gets It Wrong; this essay is about the job that fills it before the incident does.

What the manager actually does

A manager of agents does what a manager of people does, translated. Sets the limits: which decision classes the fleet may touch alone, which require a consult, which are forbidden, written down, versioned, owned. Reviews the exceptions: reads the escalations and near-misses weekly the way a good manager reads the anomalies, not the averages. Watches capability drift: an agent fleet changes when its underlying model changes, which happens silently and often, so the manager needs the behavioral record that shows their workers were swapped overnight. And answers: when the fleet gets one wrong, the manager is the name in the room, with the three tools that make answering survivable: visibility, escalation, and a kill switch that has actually been tested.

Span of control is real again

Management theory spent a century learning that one human can genuinely supervise only so many humans, and the number is small. Agents reset the arithmetic but not the principle. One human cannot meaningfully manage a thousand agents by reading their logs; they can manage them the way a plant manager runs a thousand machines, through instruments, thresholds, and alarms designed so that silence is informative. That is the real design question for the agent era: not how many agents can we run, but what instrumentation makes one accountable human per fleet honest rather than ceremonial. A manager without instruments is the crumple zone again, staffed and waiting.

Why the team owns it means nobody does

The standard objection is that the team owns the agents, which sounds robust and means, on the day it matters, that ownership is a group photo. Groups do not review exception queues; individuals do. Groups do not lose sleep over an untested kill switch; the named owner does. This is not a claim that individuals are nobler than teams. It is the oldest operational observation there is: responsibility divided by n approaches zero, and n grows with every reorg. One fleet, one name. The name can rotate. It cannot be plural.

The staffing model nobody has written

Here is the uncomfortable budget line coming for every AI-forward company: management overhead for the agent workforce. Some fraction of a qualified human per fleet, with real authority, real instruments, and consult routes to deeper experts when the fleet surfaces something beyond them, the exact seam HumanChain is built for. Companies will resist the line because agents were sold as the end of headcount. But the choice is not between free agents and managed agents. It is between paying for management by design or paying for its absence by incident, and the second invoice arrives with interest and a reporter. The agent workforce is here. Someone has to be its manager, and someone has to be able to prove it.

If you want a place to start on Monday: pick your single most consequential fleet, name its manager, and give that one person a week to answer three questions in writing: what can these agents do alone, how would I know if they went wrong, and how do I stop them. The gaps in the answers are your real agent-governance backlog, produced faster and cheaper than any framework engagement, and owned, from the first sentence, by exactly one accountable name.