The managers at Amazon’s off-site retreat had reached a sensible conclusion. Groups needed to communicate more.
Jeff Bezos got up. “No,” he said, “communication is terrible!”
That is how David Risher, a former Amazon marketing chief, remembered it when he told the story to Fast Company in 2004. The managers were shocked.
Two pizzas
What Bezos wanted, as the magazine described it, was a company where small groups could try out their ideas without waiting on everyone else. The rule of thumb that came out of it was easy to remember: if two pizzas can’t feed a team, it’s too big.
The number never settled. Accounts over the years put it at five to seven, six to ten, or under ten. People who worked in these teams say the size was never the main point. What mattered was that each team owned a product or service from end to end, with its own measure of success, and could ship without asking other teams for permission. (The case has the details.)
Fewer lines to keep up
The Harvard psychologist J. Richard Hackman was one of the leading experts on teams. As a team grows, he said, the number of links that need managing between its members goes up fast. (Brooks’s law does the arithmetic: three people share three lines, ten share forty-five.) “It’s managing the links between members that gets teams into trouble,” he told the Harvard Business Review in 2009. “My rule of thumb is no double digits. In my courses, I never allow teams of more than six students.”
A small team cuts the talk inside it. But if it still needs sign-off from three other teams before it can do anything, the talk has only moved outside. Pizza alone doesn’t fix that. Ownership does.
Two halves of one team
Imagine a marketing team of fourteen. Every Monday they spend an hour going around the table, and most of what each person hears doesn’t concern them. The manager splits them in two. Nothing changes. Both halves still need sign-off from each other on every campaign, so the Monday meeting simply becomes two meetings and a hand-off.
Then the manager draws the line differently. One group owns getting new customers to sign up for a trial, judged by how many do. The other owns keeping existing customers, judged by how many renew. Each has its own budget and can launch without asking the other. They still have lunch together. They just don’t need each other’s permission to work.
When big is better
Small isn’t always the answer. In 2019 Lingfei Wu, Dashun Wang and James Evans analyzed more than 65 million papers, patents and software products from 1954 to 2014. Small teams tended to disrupt their fields with new ideas. Large teams tended to develop existing ones, building on recent, popular work. The authors’ conclusion was that both kinds are essential.
Two pizzas fit work that needs to try things, ship and learn fast. A job that has to pull many specialties together around one well-known problem may need more chairs.
Agents that vote, agents that talk
AI agents seem at first to break the pizza rule. In 2024 a group of researchers had a model answer the same question many times and took a vote. On a set of grade-school maths problems, fifteen copies of a smaller model voting together matched one copy of a model about five times its size.
But those agents never talked to each other, and the test was one kind of maths problem. In a 2025 study, researchers from Google and elsewhere tested agents that had to coordinate, on a fixed total budget. In their setups, the rounds of reasoning and replies grew much faster than the number of agents. From that, the authors inferred that under such a budget, each agent’s share of thinking would become “prohibitively thin beyond 3–4 agents.” That is a finding about their configurations, not a universal headcount. But the mechanism would have sounded familiar to Bezos: communication crowding out the work.
So the useful question for a team of agents is close to the one Amazon asked. Can each agent have a piece of work it can finish and check against the acceptance criteria everyone agreed on? The hand-offs between pieces still need checking. Less talk is a means, not the goal.