Brooks's Law Meets AI Agents: The Bottleneck Moves
Brooks's Law survives AI as a coordination problem: agents can prototype requirements faster, but parallel work, review fatigue and unclear specs can produce the wrong answer sooner.
The Mythical Man-Month's warning that adding manpower to a late software project makes it later still applies to AI agents, though the bottleneck has moved. An agent may be productive in milliseconds, but sequential work remains sequential, and a human still has to define done, review output, catch hallucinations and stitch components together.
Without clear orchestration, nine autonomous agents on one codebase can rewrite one another's work in a high-speed loop. The author proposes a 2026 update: adding agents to a project with unclear requirements gets you the wrong answer faster.
Key takeaways
- Brooks formulated the law in 1975 from IBM's OS/360 project, where adding manpower to a late software project made it later.
- AI removes traditional bottlenecks such as onboarding time and stand-ups, but overhead shifts to managing agents: defining done, reviewing code, catching hallucinations and integrating pieces.
- The Baby Problem persists: nine agents cannot deliver a baby in one month because sequential tasks remain sequential.
- Without crystal-clear orchestration, nine autonomous agents on a single codebase will rewrite each other's work in a high-speed loop.
- The proposed 2026 update: adding agents to a project with unclear requirements gets you the wrong answer faster.
What people think
Analysis details
20 comments analyzed. Analyzed .
The sample selects active discussion branches and includes available parent comments. It may omit parts of the full thread. The analysis included 20 of 32 usable stored comments because of the input limit. Selected themes do not measure community opinion or how common a view is.
Cheap prototypes as requirements probes
A one-shotted proof of concept can expose requirements by giving stakeholders something concrete to reject or refine. The argument is that people are better at identifying what they dislike than specifying what they want, and a $4 token run may be cheaper than a $1,000-per-hour stakeholder meeting. The prototype remains a probe, not a specification, and still needs interpretation.
Speed can hide architectural drift
Fast agent-driven changes can produce a feature that misses context or a large rewrite that no one fully understands. If the developer never internalizes the problem, the wrong direction may go unnoticed. A counterargument is that cheap iteration reduces sunk-cost bias: an hour's work can be discarded without the reluctance that a week-long manual effort would create. Some also argue that determining the right thing is management's job, not the engineer's.
Parallel agents collide and distort tests
Running even two agents in parallel can cause them to step on each other, wait on each other, and make performance tests noisy because the tests run concurrently. Claims that nine coders with Claude Code can build an application quickly face the objection that the application may not actually work. A counterexample is a procedurally generated first-person shooter that worked for a clearly defined, non-contradictory spec, though writing that spec is nontrivial.
Review fatigue and uneven developer outcomes
AI amplifies existing differences: some developers become much faster, while others ship unchecked changes or stop using AI. Review fatigue becomes a major problem when work turns into mostly reviewing AI-generated code, and some teams shift toward testing and deterministic validation instead of line-by-line review. The good/bad programmer framing is disputed; fit with the workflow may matter more.
Fast feedback loops define agent strengths
Agents are most useful for grindy, trial-and-error work such as reverse engineering and performance optimization, where feedback is fast. The same models might help biology or cancer research, but physical experiments and long time horizons keep cycle times high. The useful problem space is much smaller than the one agents are marketed for.