A Unified Framework for Misunderstanding

This paper addresses the critical gap in AI safety and reliability by formalizing how misunderstandings occur, propagate, and are detected in agentic systems. By synthesizing research across pragmatics (the study of context-dependent meaning) and multi-agent systems, the authors move beyond simple 'hallucination' metrics to a more nuanced taxonomy of communicative failure. The framework categorizes misunderstandings into three distinct phases: generation (where the semantic or pragmatic intent is lost), amplification (where system feedback loops or context windows exacerbate the error), and detection (the mechanisms required to identify and correct the divergence).

Practical Application for AI Builders

The authors provide a coding manual and evidence matrix designed for immediate application in agent development. The taxonomy helps engineers move from reactive debugging to proactive architecture. By mapping specific failure modes—such as grounding failures, implicature errors, and context-drift—to their respective stages in the agentic pipeline, developers can implement targeted guardrails. The research suggests that most agentic misunderstandings are not merely 'model errors' but failures in the pragmatic alignment between the user's intent and the agent's internal state representation. The inclusion of a coding manual allows teams to audit their own agent interactions to determine if failures are rooted in linguistic ambiguity, lack of shared context, or flawed reasoning chains.