Why AI Coding Fails Without Written Contracts: Lessons from Real Breakages

AI coding assistants execute instructions literally, without questioning intent or recalling past context, which means ambiguous or unwritten agreements between system components can silently cause failures. A real-world case from a project called TradeOMS illustrates this: a single backend JSON field rename — never documented — broke 19 test files and the frontend UI simultaneously, because the response contract existed only as a chat conversation. A follow-up scan of 218 frontend API call sites uncovered 22 backend endpoints returning HTTP 500 errors, all stemming from undocumented interface drift that no human or AI had flagged. Unlike human engineers, AI models carry no memory between sessions and cannot recognize that a change conflicts with a prior agreement they were never shown. The core argument is that written, machine-readable contracts — not chat history or verbal agreements — are the only reliable way to keep AI-assisted development aligned with actual system intent.