foundations-grounding-communication
Grounding & Communication Foundations
Scope note — two senses of "grounding": This skill covers the Clark conversational sense: establishing shared meaning between parties (common ground, acceptance evidence, repair). It does not cover the LLM-attribution / RAG sense: whether a generated response is grounded in retrieved documents (citation faithfulness, hallucination detection). For the attribution sense, see ai-rag and the FACTS Grounding / RAGAS / Wallat et al. frameworks (sources FACTSGrounding2025, RAGAS2024, WallatFaithfulness2024 in data/sources.json).
10 canonical grounding-theory primitives for the process by which two or more parties establish enough shared understanding to coordinate. Founded by Herbert Clark and colleagues (1989, 1991, 1996), grounding theory is the most-cited formal account of how interlocutors solve the "do we mean the same thing?" problem efficiently.
It is the missing layer for multi-agent LLM systems: empirical work on multi-agent failures (MAST taxonomy, NeurIPS 2025) identifies system-design specification issues, inter-agent misalignment, and task-verification gaps across 1,600+ traces. Agents often proceed on different interpretations of the same brief. That is not only a coordination problem (information structure); it is a grounding problem (insufficient common ground at handoff). foundations-team-theory tells you whether agents should communicate; this skill tells you how they actually establish shared meaning when they do.
Static vs. dynamic grounding (2026). The distinction now carries the most diagnostic weight. Static grounding maps language to a shared context in one shot; dynamic grounding requires negotiating meaning across turns — joint plan formation, commitment, and execution. Yao, Zou, and Hawkins (2026) show the gap is not a reasoning-capacity problem: in an iterated negotiation game with verifiable jointly optimal outcomes, agents that identify Pareto-optimal allocations in isolation consistently fail to reach them as dyads, across models. Their four failure modes — loss of shared interaction history, anchoring to early proposals, defaulting to equal splits over reward-maximizing coordination, and referential binding errors across turns — map onto primitives #1, #3, #8, and #6 respectively. Most benchmarks and most agent handoff designs still test only the static case.