An engineer running AI agents is managing a team, starting with themselves

An engineer running three AI agents at once is doing a manager’s job, and when nobody treats it that way, a team can end up shipping more than ever while fewer of its engineers can explain how the system works. They brief each agent, check its work, decide what ships and notice when one has wandered off course, all while keeping track of their own thinking. Most were hired to build and few were trained to manage, and in that setup the hardest person to manage is the engineer themselves.

What supervising agents costs

The research is starting to catch up with what engineers describe. In a survey of 203 developers published in November 2026 by researchers at North Carolina State University and the University of Oklahoma, 64% said they often approved at least one kind of agent action without a close look. They also said their workload felt lighter, yet supervising agents stayed mentally demanding, and 63% rated at least five of ten supervision tasks as very or extremely demanding (Wu and colleagues, 2026).

Cognitive drift

I call the slide behind those numbers cognitive drift, my shorthand for something researchers have long studied. Juggling several streams of work, an engineer slowly loses hold of what they meant each agent to do and where the system now stands. They approve a change they only half follow, or lose track of what they asked for an hour ago, and each slip looks too small to matter. Human-factors researchers have studied the cost of supervising automation at least since Lisanne Bainbridge’s “Ironies of Automation” in 1983.

Margaret-Anne Storey has a name for where that leads. She calls it cognitive debt, when a team’s shared understanding of its own system comes apart, usually without a failing build to warn anyone (Storey, 2026). Drift is how that debt builds up, one engineer and one afternoon at a time, until the team as a whole can no longer explain what it built.

Coaching the manager inside each engineer

The management that matters most points inward, at the engineer’s own attention and judgment. They’re effectively managing a team while they’re managing themselves. The skills to coach for are the ones a good first-time manager learns: briefing clearly, handing off only what they can check and reviewing it against what they meant, and knowing when to stop and take the work back.

What leaders can change

Engineers shouldn’t have to work this out alone, because much of what drives drift is set by leaders. Cap how much parallel work one person carries, based on how much they can properly review, and make it a team norm so nobody has to prove themselves by running more. Measure what the business can see, like how long features take to reach customers and how often a change fails or needs an unplanned fix, and never how many agents someone runs or how many pull requests they open.

Make understanding part of what done means, too. Storey suggests that at least one person on the team should fully understand each AI-written change before it ships, and it’s a rule worth adopting. The Weekly Replay I’ve written about helps as well, because replaying a real session out loud with the team helps rebuild the shared understanding that drift wears away. Between them, they cover two of the Four Handoffs: Storey’s rule is part of Review, and the Replay is part of Learning.

If your engineers are shipping more than ever and fewer of them can explain how the system works, bring one recent change nobody could fully explain to a thirty-minute call.

Read next

Why slowing down first makes engineering faster

All writing