The Mythical Man-Month in the Age of AI
AI coding tools and autonomous agents are radically speeding up code generation, but many engineers argue that Fred Brooks’ old insight still holds: the bottleneck is coordination, understanding, and verification, not typing. Commenters describe AI as an amplifier that makes strong developers vastly more productive while enabling weaker ones to ship more unchecked “slop,” and warn that reviewing and integrating AI output can be cognitively exhausting and organizationally fragile. There is broad agreement that small, well-orchestrated teams and tightly scoped tasks benefit most, while naive attempts to parallelize agents across a shared codebase risk multiplying complexity, technical debt, and misaligned features.
Mythical Man-Month Still Applies (Mostly)
- Many argue Brooks’ core point still holds: the hard part is coordination, shared understanding, and design coherence, not typing.
- AI makes code generation cheap, but reviewing, integrating, and remembering why decisions were made are still human bottlenecks.
- Some find AI development more cognitively exhausting: faster loops, more decisions, and constant vigilance over large changes.
- Parallel agents without careful orchestration are seen as recreating the “too many cooks” problem, just faster.
Agents, Parallelism, and Orchestration
- Reports diverge: some say multiple agents step on each other’s work, create noisy tests, and rewrite each other in loops.
- Others claim large models can coordinate subagents well, using plans, git worktrees, and clear task boundaries to run many agents productively in parallel.
- There’s debate over whether using cheaper submodels actually saves tokens, since supervisor models must still re-read and review their output.
- LLMs are described as poor self-managers: overspecifying plans, locking in suboptimal designs, and rarely questioning earlier assumptions.
“Wrong Answer Faster” and Prototyping
- Many see huge value in rapidly generating bad-but-working POCs to clarify requirements and align stakeholders, compared with expensive, vague meetings.
- Benefits: tangible demos surface misunderstandings, users are better at saying what they don’t like, and sunk-cost bias is lower when code is cheap.
- Risks: flashy prototypes can seduce stakeholders into treating throwaways as shippable; POCs can hide complexity and push teams down wrong paths without enough deep thinking.
- Consensus: iteration beats one-shot perfection, but verification, testing, and human judgment remain crucial.
Productivity, Quality, and Team Dynamics
- Some report ≥10x “debugged LOC” productivity; others see increased incidents, bloat, and declining reliability and simplicity.
- AI is widely described as an amplifier: strong, motivated devs become dramatically more effective; weak or disengaged devs ship more unvetted slop faster.
- Review fatigue is a major pain point; engineers feel reduced to PR reviewers and QA, which many find demotivating.
- Single engineers plus agents or very small teams are viewed as increasingly optimal; large teams with AI risk coordination overload.
Cognitive and Cultural Concerns
- Some enjoy the “engineering director” role—specifying, critiquing, refining—more than line-by-line coding.
- Others find agent interaction tedious, worry it erodes their reasoning and craftsmanship, and prefer writing code themselves.
- Several worry that organizational incentives (velocity over quality) plus AI will accelerate tech debt and “shadow IT” rather than durable systems.