Understanding decay
The common claim
If AI writes the code, you stop holding the system in your head. Six months later nobody can change it cold — including you.
What I actually believe
Nobody holds a large codebase in their head past a certain size. “Bus factor = one human’s memory” was always a myth.
This is not vibe coding. Intent and architecture guidance go in. I expect the same from the team. The human job moved up a level: how it fits together, invariants, boundaries — not line-diff archaeology across twenty files.
Debugging is asking good questions with tools. AI plus logs plus Sentry is real leverage. Often you don’t need the whole map — you need the path to this failure.
Where the architecture lives
Greenfield: an architecture spec that steers the agent.
Existing systems: introspect carefully and extract a documented architecture. We’ve done that in detail on a current system (Java / Spring Boot 3, React and TypeScript). Elsewhere I haven’t gone as far as I’d like — mostly bandwidth, not disagreement with the idea.
Line-by-line PR archaeology doesn’t scale. Architecture fidelity does.
Where it still bites
Bug-finding is not design judgment. Instrumentation tells you what blew up. It doesn’t invent the right boundary for auth, money, or idempotency before the next feature ships.
Cold moments without tools — whiteboard, interview, outage when the agent or the logs are dark — still need a human who can draw the decision map. That’s a soft spot to own, not a reason to pretend AI-native delivery is unserious.
Drafted with AI. Reviewed and edited by Ed Henderson.