Multi-Agent Failure Behaviors

This book started with a bug report that couldn't be filed anywhere, because nothing had technically failed. Eight subagents, each assigned a chapter of an audit plan, were asked to review their own chapter. All eight answered a different question than the one asked, in precisely the same wrong way. No exception, no malformed output. Just fluent, professional, wrong.

Multi-Agent Failure Behaviors is a field guide to the systematic ways LLM-based agent systems go wrong, organized into five families: Role Contamination, Underspecification, Judgment Bias, Chained Multi-Agent System Pathologies, and Goal Misalignment. Every failure behavior gets the same treatment: a plain-language definition, a narrative example that makes the mechanism intuitive, a minimal TypeScript implementation that reproduces the failure, a corrected implementation with the design principle spelled out, and a diagnostic probe you can run in CI.

Most hallucinations in agentic pipelines are not invented at a single point; they are manufactured by the system. An agent fills a gap with a plausible guess, a judge waves it through, the next step strips the hedging, and three steps later it is an established fact in an executive summary. This book shows where that process can be interrupted, with defenses that do not depend on the model's cooperation: separated identities, disjoint contexts, structured data at boundaries, least privilege, mechanical enforcement.

Written for engineers and architects building agentic systems who want a vocabulary and a checklist for failures that do not look like failures. Code samples are TypeScript against a standard LLM messages API; the mechanisms are model-agnostic.

The author is a CTO, AWS AI Hero, and Cursor Ambassador who has worked on neural networks since his 2010 PhD and on production agentic systems since they existed. He observed, diagnosed, and fixed every failure behavior in this book in real systems before it had a name.