Wiki / concepts / wiki

Feedback-loop ceiling

"The quality of your feedback loops influences how good your AI can code. That is the ceiling." โ€” matt-pocock. Without tests/typechecks the AI codes blind; improving loop quality beats blaming the model.

How to raise the ceiling:

  • TDD (red-green-refactor) โ€” "absolutely essential"; instrumenting code before writing it makes test-cheating structurally harder than writing implementation-then-tests in layers.
  • deep-modules (Ousterhout) โ€” small interface, big functionality, one big test boundary; unwatched AI produces shallow-module sprawl that can't be cleanly test-bounded. Pocock's browser video editor wrapped front-to-back in one outside-testable module was "night and day" for AI effectiveness.
  • tracer-bullet-slices โ€” vertical slices make feedback arrive in phase 1 instead of after all layers land; AI naturally codes horizontally.
  • Gray boxes โ€” design the interfaces, delegate the implementations: keeps the human's mental model of the codebase without reviewing every line.

Pocock's corollary: "bad code bases make bad agents." The architectural echo of dhh's Basecamp-5 lesson in vibe-coding (unreviewed agent PRs destroyed the architecture) โ€” and the constructive counterpart to ai-removes-the-maintenance-tax.

Source: report

Pipeline version (zazencodes, Nov 2026): in a software-factory, a deterministic gate runs the project's lint, type check, tests and build after every build round. A failure goes straight back to the builder with the log, so no reviewer agent spends tokens on code that doesn't pass (report).

Linked from

Deterministic spineMatt PocockModels need tokens to think