What Managers Run Out Of
The pitch says one manager can now run twenty teams. The org chart believes it.
The pitch keeps showing up wearing different clothes. Automate enough of the coordination work and one manager can effectively run twenty teams instead of five. Agents don’t need one-on-ones, don’t burn out, don’t need a Friday afternoon to recover. Widen the span, remove a layer of management, and the org chart gets flatter and cheaper at the same time.
Igor Rikalo, president and COO of supply-chain software vendor o9 Solutions, put a number on where this goes in a July 2026 Forbes piece. Within five years, he expects a strong product manager to be coordinating something like 180 agents, and a head of supply chain something like 300, running forecasts, negotiating contracts, testing scenarios, and flagging risks in real time. That’s one vendor executive’s forward projection, not a measured outcome anywhere yet, and it names the assumption plainly. Swap the direct report for an agent, and span of control turns into mostly a capacity problem where adding compute means supervising more work.
Twenty, Not Two Hundred
Meta’s own numbers are nowhere near 180 or 300, but the company still ran a smaller-scale version of the same wider-span bet this year, and not for the reason the pitch assumes. Its new Applied AI division, formed this spring to consolidate generative AI work, left a number of managers with well over twenty direct reports each. By June, CTO Andrew Bosworth was telling employees that morale was among the worst he’d seen in twenty years at the company, and committing in a memo to cap managers at roughly twenty reports going forward, limit how often people get reassigned during restructurings, and offer employees optional AI coaching tools if they wanted them.
Notice what’s doing the work in that fix. It isn’t better agents. It’s fewer people per manager and more of the manager’s actual attention per person. The AI coaching tool is a minor offering on the side, aimed at employees rather than the managers carrying the wider spans, and it isn’t the mechanism that’s supposed to make twenty tolerable where the pre-cap ratio wasn’t.
Jade Rubick, who coaches engineering leaders on exactly this tradeoff, put his finger on the same gap in a piece published days before Rikalo’s. A lot of the current push toward wider spans, he argues, is really an old anti-management bias wearing an AI justification, where companies want fewer layers and hope the tooling covers for the coordination they’re removing, without doing the groundwork to check whether it actually does.
What Doesn’t Get Smaller
This connects to something I’ve written about before at the level of a single delegated task. Span of control is that same gap multiplied across an entire team, and the org-chart math skips it entirely. Managing twenty people was never mostly a scheduling problem. It was deciding whose work needed a second look this week, whose one-on-one couldn’t slip again, which disagreement needed you in the room and which one would resolve itself. That’s interpretive work, and it scales with judgment calls made, not with reports counted.
Swap five of those direct reports for agents and the judgment calls don’t disappear. They change shape. An agent’s output that nobody reviewed isn’t automatically wrong, but confidence in it is ungrounded without knowing why it made the calls it made, and deciding which outputs earn that scrutiny this week is the same interpretive work a manager already does for human reports, just arriving faster and in higher volume.
A new report’s design decision takes days to land on your desk, with plenty of natural checkpoints along the way. An agent’s takes minutes. The risk isn’t that agents introduce a new kind of failure. It’s that they remove the pacing that used to force the review to happen.
Even granting that any single agent needs less attention than any single person, a manager overseeing 180 of them isn’t reviewing 180 miniature versions of one report’s workload. They’re fielding whatever share of those 180 outputs actually needed a human this week, arriving on no predictable schedule at all, which is a different and harder job than the org chart’s division problem implies.
Aviation decision engines taught me a version of this years before agents were the vendor pitch. A recommendation engine that suggests a routing change can be technically correct and still need a human to ask why, because the cost of trusting a bad one without asking isn’t a missed deadline. Widening span of control just multiplies how many of those “why” questions are waiting for you at once, whether the report typing the answer is a person or a model.
That’s the piece I keep watching for in my own teams, human and AI-assisted both. The question isn’t whether the work is getting done. It’s whether I’ve actually built in a moment to ask which of it deserves a harder look before it ships, and that’s an easy thing to skip when the throughput looks fine.
The Corrected Math
Meta’s twenty isn’t a magic number, and neither is Rikalo’s 180 or 300. But Meta’s retreat is what happens when the org-chart-variable version of span of control meets an actual team, and Rikalo’s projection is the same assumption stated at ten times the scale, before anyone’s tested it there either. Adding agents to one side of the equation doesn’t change what’s actually being managed on the other. The judgment work still has to happen somewhere, by someone, on a schedule that lets it happen well.
Before adding another direct report, human or agent, to anyone’s plate, ask what specific time for review and judgment comes with it, not just what capacity gets freed up. If the answer is “the AI will handle that too,” that’s the assumption worth testing before the org chart finds out the hard way.
