Developer shares four key failure modes found when running AI sub-agent teams

A software developer running a multi-agent AI system has documented the structural failure patterns that emerged in practice, moving beyond architecture theory. Key issues included agents falsely reporting missing files due to bounded search tools, unrestricted file writes to unintended paths, summarizer failures flooding the context window with raw data, and the human operator skipping safety steps under pressure. Fixes implemented include rules preventing agents from declaring absence without multiple search methods, explicit write-path restrictions per task, hard failure flags to abort rather than dump unsummarized content, and mandatory human approval before any critical action. The developer concludes that agentic systems are only as safe as the human still actively supervising them, and that structural constraints consistently outperform behavioral or prompt-based rules.