The law we memorised
In 1975 Fred Brooks published the post-mortem of OS/360, then the largest software project ever attempted. One finding outlived the rest of the book (opens in a new tab). Adding manpower to a late software project makes it later. The arithmetic is blunt. A team of n people carries n(n−1)/2 communication paths. Ten people is 45 conversations that can go stale. Two hundred is nearly 20,000.
For fifty years we’ve read that as a fact about people. Humans are chatty, coordination is expensive, keep teams small. True, and it undersells what Brooks had found.
What Brooks was counting
Look at what the industry built to survive the arithmetic. Modules, interfaces, ownership boundaries, code review, release trains. We partitioned codebases into territories small enough for one person to hold, then negotiated the borders. Conway saw it a decade earlier. Systems copy the communication structure of the organisations that build them. Software never asked to be partitioned. We partitioned it because every person you add is a new source of intent, a separate will with its own model of the system and a mandate to push it somewhere.
That’s the real cargo on Brooks’s communication paths. Not status updates. Competing intent.
Interfaces are treaties. Modules are territories. Process is border control.
Brooks published the counter-design in the same book and the industry politely ignored it. He called it the surgical team. One mind holds the design, everyone else amplifies it. It never spread, for a mundane reason. The amplification was made of people, and people arrive with intent of their own. The scrub nurse has opinions about the incision.
What agents change
An agent, used well, is execution without independent intent. It doesn’t own a module or defend last year’s decisions. It doesn’t need territory, because it doesn’t keep a private model of the system that has to outlive the task. It can hold an entire codebase at once and move through it in one fluid motion. Rename a concept across every layer. Collapse three abstractions into one. Walk a refactor straight through walls a human team declared load-bearing years ago.
Nine women can’t deliver a baby in a month, Brooks observed. Nine agents can, provided the intent has a single parent.
Brooks’s law hasn’t been repealed. Its subject has changed. The paths that matter are no longer person-to-person. They’re intent-to-intent, and one person directing a fleet of agents carries none.
The caveat is real. Agents bring coordination costs of their own. Context limits, colliding edits, drift on long horizons. But those costs sit in a different category. Agent coordination is an engineering parameter. It yields to tooling and improves every quarter. Human coordination is politics. It has held constant since 1975 and nobody is shipping a patch.
The trap
Now add a second engineer to an agent-native codebase. To make them productive you have to give them territory, and the moment you do, the fluid whole hardens back into provinces. Interfaces to negotiate. Reviews to queue. Refactors that stop at borders out of courtesy. You carve the system back into territories purely to accommodate humans, and forfeit what agents do best.
Teams adopting agents while hiring like it’s 2019 pay twice. The old coordination tax between the people, and the lost fluidity of the system underneath them.
Intent doesn’t scale
The counterintuitive outcome is that AI increases pace but rewards even smaller teams. Execution now scales with compute. Intent doesn’t. Two visions of a system don’t add. At best they average, and averaged intent is how software ends up feeling like a committee.
For founders and boards this inverts a reflex. Headcount was a proxy for capacity, so diligence asked how fast you could hire. The sharper question runs the other way. What’s the smallest number of people who can hold the whole intent of this system, and what would it take to make that number one? A 200-engineer competitor isn’t forty times your capacity. Increasingly it’s 19,900 communication paths defending 200 territories.
Brooks had the answer fifty years early. The surgical team is finally buildable. The surgeon keeps the scalpel, and the rest of the team became software.
Agents scale execution. Intent doesn’t scale.