Do LLM agents misinterpret delegated requests more than formal protocols?

The paper contrasts two ways multi-agent systems establish shared understanding of a message. Classical AAMAS communication languages such as KQML and FIPA-ACL "define not only message syntax but also the semantics of communicative acts like requests, commitments, or promises," so that when one agent sends a message, both sides have a formally guaranteed interpretation of what was asked, committed, or promised. Most agentic AI systems instead "rely on unstructured natural language dialogue, mediated by large language models," which the paper says is "flexible and expressive" but "lack[s] guaranteed semantics: meaning is inferred statistically rather than interpreted formally."

The paper illustrates the stakes with a scenario: someone wants to know when a guest arrives tomorrow, in order to avoid them, and asks their personal agent to query the host's agent. With a formal protocol, "I just want that agent to state the arrival time if it can look it up in its own information" — a strictly scoped request for information. Without one, a natural-language agent "might start asking the guest for the arrival time on my behalf (giving my request as the reason for the question)," overstepping the delegated task and creating an "awkward situation" the user never asked for. The point is that formal speech-act types let a sender distinguish "ask the host" from "ask anyone who might know," a distinction natural-language dialogue does not enforce, which the paper says "can lead to misunderstanding, inconsistent commitments, or spurious alignment between interacting agents."

This gives a concrete coordination-layer version of the structural argument in Can classical agent architectures fix agentic AI accountability?, where the general claim that LLM agents lack formal commitment semantics is shown here as a specific delegation failure. It also qualifies Can workplace AI risks emerge from interactions alone?: that note models interaction risk as emergent from well-functioning parts, and this excerpt supplies one concrete mechanism for such emergence — an ambiguous natural-language request that gets silently over-executed. And it touches How can users verify what an AI agent actually plans to do? from the other direction: formal communicative-act semantics build common ground into the protocol itself, rather than relying on after-the-fact verification of what an agent did.

The paper offers no measurement of how often natural-language agent-to-agent communication actually produces overreach of this kind in deployed systems; the arrival-time case is a hypothetical illustration, not an observed incident or benchmark result. Nor does it specify how a constrained communication layer would be retrofitted onto LLM-based agents without losing the flexibility that motivated moving away from KQML/FIPA-ACL in the first place. The implication the excerpt supports is narrower than "agents need formal protocols" — it is that unconstrained natural-language delegation has at least one identifiable, structural failure mode that formal speech-act typing was designed to prevent.