Engineering leadership · Book notes
The Mythical Man-Month asks whether agents count as manpower
Agents skip the ramp-up that makes Brooks's law bite, but do they dodge it or just move the communication cost onto the humans who review and release?
When programs are late, you slip the date and add more headcount. In The Mythical Man-Month (1975), Frederick Brooks wrote that adding more people to a software project that is already late will make it even later. The 1995 anniversary edition includes the original text, as well as “No Silver Bullet” and a twenty-year retrospective.
Martin Fowler revisited the book this month and said the most valuable lesson for him was conceptual integrity. The reflex has a new form: a widely shared ExperiencedDevs thread described how managers assumed a 5x productivity increase from AI agents before work even started, turning a five-person, two-month plan into twelve days.
Brooks’s law has two costs: the time newcomers need to ramp up before they can contribute, and communication paths that grow as n(n-1)/2 (so four people need six channels, seven need twenty-one). The line that carries the argument in chapter 2 is “Men and months are interchangeable commodities only when a task can be partitioned among many workers with no communication among them.”
The law still holds for human teams. A recent ExperiencedDevs post described a team that was three weeks behind, added three developers, and finished five weeks behind. The original four spent about half their time explaining the rationale behind past decisions. No review tool could tell the newcomers why the auth layer had the shape it did.
Agents seem to avoid the first cost: they don’t ramp up like humans, and you don’t need to call an onboarding meeting. So the law seems optional. But Google Research measured multi-agent setups and found coordination improved by ~81% on work that could be cleanly split, but got 39 to 70% worse on sequential work. That’s Brooks’s partition condition, measured on silicon. And when agents operate independently without an orchestrator, errors get amplified up to 17 times.
The book’s other ideas matter more for large programs. Conceptual integrity means the design comes from one mind or a small group in agreement. A surgical team has one lead and supporting specialists, and the Hacker News discussion of Fowler’s post reframed it in terms of one person directing agents playing multiple roles. In regulated shops, the sequential delivery stages (architecture review, testing, security sign-off, change approval) don’t magically get faster just because the code got cheaper.
The DORA 2025 report says AI adoption raises throughput but delivery stability still drops, and the gains go to loosely coupled systems. McKinsey and Oxford say large IT programs already run 45% over budget and deliver 56% less value than planned. Modernization and bank transformation timelines.. do agents change the failing number, or just add another vendor stream to coordinate? Some vendor pitches promise decade-long mainframe programs in three years.
I’ve spent the last ten years working with large financial institutions and Fortune 500 companies, now as a principal cloud architect. Before that I was a developer, then a development manager. If agents are capacity, then the things Brooks protected become design work.. a small group holding the design, clean partitions, fewer channels. The human reviewers and approvers become the scarce channel all agent output must pass through. So when a program is late, do you add more agents? Or do you first reduce the number of things that have to talk to each other?