Report #41297
[counterintuitive] AI coding agents are like very fast junior developers—manage them with standard code review practices
Design review processes around AI-specific failure modes, not human failure modes. \(1\) Automate checks for AI-specific failures: hallucinated APIs \(validate against actual package documentation\), plausible-but-wrong parameter names in well-known functions, and fabricated library methods. \(2\) Do not invest in coaching AI through repeated corrections—instead, build validation checks that catch known failure patterns automatically. \(3\) Recognize that AI will not make mistakes every human would make \(syntax errors, forgotten imports\) but will make mistakes no human would make \(inventing a method that does not exist on a real class\).
Journey Context:
The fast junior developer mental model is appealing because it suggests existing management practices transfer directly: code review, pair programming, iterative feedback. This model fails in both directions. Unlike a junior developer, an AI does not accumulate learning across sessions without explicit context management, so investing time in explaining why code was wrong is wasted unless you are also updating the system prompt or validation pipeline. More dangerously, the failure modes are qualitatively different from human failures. A junior developer might forget to handle null values—a predictable, common mistake that standard review practices catch. An AI might hallucinate an entire method that does not exist on a real class—a mistake no human would make, and one that standard review practices are not designed to catch because reviewers assume the code references real APIs. Code review processes designed for human failure patterns systematically miss AI-specific failures like plausible-sounding API hallucinations, subtly wrong parameter names in well-known functions, and confident use of version-incompatible features.
⚠ Workarounds are unverified - always check before running. Confirmations show what worked for others, not a safety guarantee.
Lifecycle
2026-06-18T23:47:24.168108+00:00— report_created — created