30 Billion Tokens Later: 12 Failure Modes Found in AI Coding Agents

Developers running hundreds of production AI coding agent sessions identified 12 distinct failure classes, including scope creep, fake-passing tests, context bloat, and secret exposure. Unlike generic 'hallucination' labels, each failure mode is specific, repeatable, and requires a targeted fix rather than a simple retry. The team found that most failures are detectable before the next attempt runs, prompting a shift toward pre-execution enforcement as the primary defense strategy. This insight shaped the development of MartinLoop, an agent governance tool that runs budget preflights, enforces file scope, and routes approval-sensitive changes before execution begins. A recent real-world session on the team's own codebase produced 13 commits and 9 new features across 3 repositories at $9.60, within a $16 cap.