I Let AI Plan 170 Changes. It Made the Same 3 Mistakes Every Time.

Developer Debashish Ghosal tested an AI-driven planning engine against 170 real-world change goals to evaluate its reliability. The study revealed that regardless of the model size, the AI consistently failed in three specific areas: unverified dependencies, unsafe sequencing, and weak rollback mechanisms. Ghosal argues that these are structural planning failures rather than limitations of model intelligence, meaning they cannot be solved by simply using larger LLMs. Instead, the author proposes a hybrid approach using deterministic gates and a 'precondition closer' to validate plans before execution. By implementing these code-based safeguards, the engine successfully shifted from merely diagnosing bugs to acting as a robust regression gate. The findings suggest that for complex multi-step tasks, developers should prioritize deterministic validation over relying solely on LLM reasoning, as structural integrity is best enforced through code rather than prompts.