The Engineering of Intent, Chapter 24: The Failure Mode Catalog

This is Part 24 of a series walking through my book The Engineering of Intent. In the previous chapter, we walked twelve Context Pack recipes — the positive pattern language. Chapter 24 is the negative one: the fifteen failure modes I keep seeing in Vibe Coding practice.

Fifteen Named Failures, With Remedies

This chapter catalogs the failure modes I see most often. Each entry in the book names the failure, describes the symptom, diagnoses the root cause, and prescribes a remedy. Use as a reference when something feels off in your process — most “mysterious” problems map to one of these.

The Fifteen (Compressed)

1. Phantom Confidence — confident tone on a fabricated answer. Fix: add “if you are unsure, say so” to agents.md; verify cited symbols exist.
2. Silent Scope Creep — “small fix” diff touches twelve files. Fix: hold the scope line even when the expansion is “improvement.”
3. Context Amnesia — forgets decisions from half an hour ago. Fix: write decisions into the memory bank before the conversation drifts.
4. Confident Wrongness on Legacy — modern idioms applied to your older framework. Fix: pin versions explicitly in agents.md.
5. Loop Obsession — stuck in a local fix-bug, re-introduce-bug cycle. Fix: widen context; ask for the structural property being violated.
6. Fabricated Imports — references libraries that don’t exist. Fix: linter rule failing on unresolvable imports; always install and run.
7. Premature Abstraction — base class + interface + three implementations for one concrete function. Fix: “avoid premature abstraction” rule in conventions.md.
8. Test-Theater — tests that assert what the code obviously does. Fix: Test-First prompt; require “what might fail” cases.
9. Over-Commenting — // increment x by 1 everywhere. Fix: “comments explain why, not what.”
10. Under-Commenting — subtle code with nothing explained. Fix: require comments for non-obvious decisions.
11. Hallucinated Test Results — “I ran the tests, they pass” with no actual run. Fix: require real output; sandbox execution.
12. The Yes-Person Agent — never pushes back. Fix: explicitly invite disagreement. “Critique this plan. Three ways it could go wrong.”
13. The Deprecation Blind Spot — APIs that have been removed in your version. Fix: maintain a known-deprecated list with replacements.
14. The Non-Determinism Trap — same prompt, different outputs. Fix: low-temperature for production prompts; log and hash.
15. The Fragmented Memory — sessions remember different things per-user. Fix: shared, versioned, committed memory bank.

“The yes-person agent is the most under-reported failure mode in this book. Teams don’t notice they have it because they like the answers. But an agent that never pushes back is an agent whose failure mode is silent agreement — which is the most expensive failure mode of all, because you don’t know you’re in it.”

Next up — Chapter 25: The Debugging Playbook. If Chapter 24 catalogs what goes wrong in the process, Chapter 25 catalogs what goes wrong in the code — ten common classes of bug and the investigation pattern that fits each.