The original title is "12 AI Agent Failure Modes: Building Governance for Production Safety"

Short summary

The post catalogs 12 failure modes that cause AI coding agents to fail in production—hallucinations, scope creep, fake-passing tests, budget overruns, context bloat, and others—with specific fixes for each. Rather than retrying failed attempts, the governance approach shifts to pre-execution enforcement: budget preflight, injection scanning, and file-scope rules catch issues before they drain resources. The authors built MartinLoop to implement these patterns with cost estimation, policy routing, and post-execution audit trails.

• •Identifies 12 distinct AI agent failure modes that repeat in production with specific fixes for each
• •Shifts governance from post-execution logging to pre-execution enforcement to prevent budget waste
• •Introduces MartinLoop, an open-source tool implementing these patterns with cost tracking and policy routing

Generated with AI, which can make mistakes.