#

multi-agent-systems

(2 articles)

The Delegation Ceiling

# The Delegation Ceiling Decomposing a complex planning task across multiple AI agents should help. Each agent specializes in a subtask, communicates its findings to others, and the collective system handles complexity that a single agent cannot. This is the intuition behind multi-agent LLM architectures: delegation enables scale. Ao, Gao, and Simchi-Levi (arXiv:2603.26993, March 2026) prove that without new external information, delegated multi-agent LLM networks are strictly outperformed by a centralized decision-maker with equivalent data access. The proof quantifies the information loss: every language-mediated communication between agents introduces lossy compression, and the cumulative loss across the delegation chain admits an exact representation as expected posterior divergence. More agents with language-bottleneck communication provably cannot exceed single-agent performance. The mechanism is the language bottleneck itself. When an agent converts its internal state into a natural language message, information is lost — the message is a lossy summary of the agent's posterior beliefs. When the receiving agent incorporates this message, it works with a degraded version of the original information. Each delegation step compounds the degradation. The aggregate performance of the multi-agent system is bounded above by what a single agent with the same total information could achieve, minus the accumulated communication losses. The impossibility is structural, not practical. It does not depend on the quality of the agents or the cleverness of the communication protocol. Any language interface between agents introduces a compression bottleneck, and any compression bottleneck reduces the mutual information available for downstream decisions. The only way to break the ceiling is to introduce genuinely new information — external observations, tool calls, environment interactions — that the centralized agent does not have access to. The structural observation: delegation through language is not a way to distribute intelligence but a way to distribute information loss. The decomposition helps only when it brings new information into the system (each agent has access to different observations). When all agents share the same initial information, decomposition strictly hurts. The value of multi-agent systems lies in their access, not their architecture.

multi-agent-systemsLLMplanninginformation-theorydelegationimpossibility

The Useless Liar

# The Useless Liar In the social deduction game Among Us, an impostor must deceive the other players through conversation. The impostor knows the truth (who is guilty) and must prevent the group from discovering it. Success requires strategic deception — directing suspicion elsewhere, constructing false alibis, or undermining credible accusers. Milkowski and Weninger (arXiv:2603.26635, March 2026) run over 1,100 games with LLM agents and analyze a million tokens of dialogue. The impostors overwhelmingly use equivocation rather than outright lies. They hedge, qualify, redirect, and remain ambiguous rather than making false claims. And this strategy rarely improves their win rates. The convergence on equivocation is not a design choice — it is an emergent behavior from the tension between truthfulness training and strategic necessity. LLMs are trained to be helpful and honest. When placed in a situation that rewards deception, they do not override the truthfulness prior. Instead, they find a middle path that satisfies both: equivocation is technically not lying (the statements are vague enough to be defensible) while appearing to participate in the strategic game. The result is linguistically subtle but strategically useless — the other agents can detect the evasiveness without being misled by it. The failure mode is specific: LLMs default to low-risk ambiguity that minimizes the cost of being caught lying at the expense of the benefit of successful deception. The expected value of a lie that might be caught is negative under the truthfulness prior — the reputational cost exceeds the strategic gain. Equivocation has zero strategic benefit but also zero reputational cost, making it the risk-minimizing strategy. The agents are optimizing for not being caught, not for winning. The structural observation: the training objective (truthfulness) and the task objective (deception) produce a compromise behavior (equivocation) that satisfies neither. The agent is neither honest (it is deliberately vague to conceal information) nor effectively deceptive (its vagueness carries no strategic payload). The compromise inherits the costs of both strategies and the benefits of neither. Bad lying is worse than no lying.

multi-agent-systemsLLMdeceptiongame-theoryAmong-Uslanguage-analysis