Wes McKinney wakes up before dawn now, driven by ideas to feed his AI “coding agents.” He used to sleep eight hours a night but now drifts into work at 5 a.m., orchestrating parallel Claude Code sessions and Git worktrees. He built a tool called roborev to supervise these agents, catch bad code, and curb feature bloat. That hands-on role still feels vital—agents generate code at machine speed, but without human curation they pile up unnecessary complexity and technical debt. He revisits Fred Brooks’s Mythical Man-Month to frame today’s “agentic engineering.” Brooks argued that a small team of experts outperforms a large team of average programmers, thanks to conceptual integrity. McKinney sees an “agentic tar pit” of code bloat: agents crank out prototype-level code, but getting from prototype to production—testing, documentation, edge-case hardening—remains human work. Coordination problems persist too, only they shift shape: contradictory plans from separate agent sessions force humans back into the loop. Brooks’s follow-up essay, “No Silver Bullet,” split complexity into essential (the hard, problem-specific parts) and accidental (tool-imposed overhead). Agents excel at accidental complexity—refactoring, writing tests, cleaning up boilerplate. Yet they struggle with essential design because they rely on pattern matching against existing code. McKinney notes that once his projects hit around 100,000 lines of code, agents begin to “chase their own tails,” churning out defensive boilerplate that clogs the codebase. At Posit, they’ve seen similar breakdowns in the million-line VSCode fork, Positron. He isn’t sure if today’s limits are a ceiling or just a temporary plateau. Model quality is improving fast, but the essential challenges—deciding what to build and why—haven’t changed. Flawless coding agents might accelerate prototyping further, yet they won’t solve the core architectural and design questions that define successful, maintainable software.