In July 2001, as a book by three of its leaders later told it, Google was nearly three years old and had several hundred employees. Its head of engineering, Wayne Rosing, was unhappy with his managers. They were strong engineers, he felt, but not great managers.
So he, Larry Page and Sergey Brin tried something radical. They got rid of every manager in the engineering organization. All the software engineers would report directly to Rosing. Rosing and Eric Schmidt gave the new structure a name: the “disorg.”
The founders loved it. They had come from a university, where students gather around projects and nobody is “managed.” Why put a manager between the company and the engineer doing the work?
The executive coach Bill Campbell, who had just started working with the company, didn’t buy it. One evening, according to the book his Google colleagues later wrote about him, he walked down the hall with Page and Brin and asked an engineer whether he wanted a manager.
Yes, the engineer said. “I want someone I can learn from, and someone to break ties.”
The questions that pile up
Google ran without engineering managers for more than a year, by that book’s account, and brought them back near the end of 2002. In Harvard Business Review’s telling, the reason was that too many people went straight to Page with “questions about expense reports, interpersonal conflicts, and other nitty-gritty issues.”
Even when no single question is hard, enough of them together will jam the one door they all queue at.
In 1933 a management consultant named V. A. Graicunas tried to put a number on that queue. A manager, he pointed out, doesn’t only deal with each person. He also has to keep in mind how they deal with each other, in pairs and in groups. Counting all of those, a manager with four people has 44 possible relationships to watch. Add a fifth person and it becomes 100. These are relationships his model says could arise, not a measurement of how busy any manager actually is. Still, the count grows much faster than the team. Outside routine work, he concluded, a manager should have no more than five direct reports, and probably only four.
Four or thirty
Here is the puzzle. Google eventually concluded that managers do matter. Its Project Oxygen research team, which presented its first questions in early 2009, set out to settle the matter. “We actually ended up trying to prove the opposite case, that managers don’t matter,” its colead Neal Patel told HBR. “Luckily, we failed.” Yet the same article notes that it is “not uncommon” at Google “to find engineering managers with 30 direct reports.”
Thirty is six times Graicunas’s limit. Both can be right, because Graicunas’s count assumes the people all need each other, and the manager has to keep track of all of it. He quoted an older rule of thumb from the British general Sir Ian Hamilton: “the smaller the responsibility of the group member, the larger may be the number of the group.” Engineers who own their work and settle most things among themselves need a manager for teaching and tie-breaks. A team where every task touches every other needs one for much more.
Most American managers sit well below thirty. Gallup’s 2025 data put the average number of people reporting to a manager at 12.1, nearly 50% higher than when it first measured in 2013. The median stayed at about five or six. A minority of very large teams pulls the average up.
Picture a manager who has just inherited a second team because a colleague’s role was cut. Fourteen people now. Her calendar fills with one-on-ones, her inbox with small approvals. Nobody has asked her anything difficult. She just can’t get to all of it before the questions start piling up at her door, as they once did at Page’s.
How many agents to send
An AI agent that hands work to other agents has a span-of-control problem too. Its limit isn’t time for one-on-ones. It is judging how many helpers a job needs and telling each one what to do.
When Anthropic built its research system, a lead agent split each question among subagents. Early versions misjudged the size of the job. Among the errors the engineers list: “spawning 50 subagents for simple queries.”
So they wrote rules of thumb for staffing into the prompts. “Simple fact-finding requires just 1 agent with 3-10 tool calls, direct comparisons might need 2-4 subagents with 10-15 calls each, and complex research might use more than 10 subagents with clearly divided responsibilities.”
More than ten helpers, by those rules of thumb, only once each one knows exactly which piece is its own.