Source: Simon Willison's Weblog — 2026-08-19
Summary
Simon Willison revisits Fred Brooks' "conceptual integrity" from The Mythical Man-Month — the idea that well-designed software feels coherent, covers the right domain, and has no surprises — and argues it's under more strain than ever with coding agents in the loop. Agents make it trivially easy to bolt on features fast, which Willison compares to the Winchester Mystery House: a 160-room California mansion that kept growing in odd directions for decades because building never stopped long enough for anyone to ask whether the whole still made sense.
Key Takeaways
- The bottleneck with agentic coding isn't typing speed or raw output volume anymore — it's a human's capacity to hold the whole system's design coherent in their head while code gets generated faster than anyone can review it for fit.
- Agents don't have an instinct for conceptual integrity; they'll happily implement a locally reasonable feature that quietly contradicts the system's overall design, because "does this fit the whole" isn't something a per-task prompt naturally captures.
- The Winchester Mystery House analogy works because that house wasn't built badly room-by-room — each room was fine on its own — it just never had a moment where someone asked whether the accumulated whole still made sense.
- The practical implication is that code review needs to shift toward auditing for architectural coherence, not just correctness of individual generated changes, since agents will happily generate code that passes tests while quietly eroding the design.