Can LLM agent groups reliably reach consensus together?

Multi-agent LLM systems are increasingly deployed in contexts that require consensus: agreeing on a delegated task, validating a shared decision, converging on a planned action sequence. The question is whether they can actually reach agreement reliably when challenged.

"Can AI Agents Agree?" (2603.01213) tests this with a Byzantine consensus game over scalar values using synchronous all-to-all simulation. The setup deliberately strips out value-optimization concerns: in a no-stake setting, agents have no preferences over the final value, so the evaluation focuses purely on agreement reachability rather than on what gets agreed to. The simplest possible test of consensus capability.

The finding is uncomfortable for current MAS deployments: valid agreement is not reliable even in benign settings without Byzantine agents, and degrades monotonically as group size grows. Introducing even a small number of Byzantine agents further reduces success. Across hundreds of simulations spanning model sizes, group sizes, and Byzantine fractions, the LLM-agent groups frequently fail to reach valid consensus within the round limit.

The mechanism is the key insight. Failures are dominated by liveness loss — timeouts and stalled convergence — rather than by subtle value corruption. The agents don't get tricked into the wrong answer; they get stuck not converging on any answer at all. This contrasts with the standard intuition that Byzantine fault tolerance is primarily about defending against adversarial value injection. For LLM agents, the harder problem is reaching agreement at all, before even worrying about whether the agreement is the right one.

The structural diagnosis: current LLM agents lack the protocol discipline that distributed systems achieve through deterministic state machines. Each agent generates stochastic responses, can drift off-topic, can fail to recognize when consensus has been reached, can introduce procedural confusion that prevents the round-limit from terminating productively. Liveness — the property that the system eventually decides something — is harder than safety (the property that what it decides is correct) when the agents themselves are stochastic.

This connects to Why do multi-agent LLM systems converge without genuine deliberation? from a different angle. Silent agreement is convergence-too-early on a wrong answer; this paper documents the opposite failure mode — failure-to-converge-at-all. The two together bracket the consensus failure space: when MAS systems try to reach agreement, they either (a) prematurely silently agree without genuine deliberation, or (b) fail to converge through liveness loss. Neither is reliable.

The implication for deployment is stark: agreement is not yet a dependable emergent capability of current LLM-agent groups. Systems that rely on multi-agent consensus for cooperation, delegation, or safety-critical coordination are building on a fragile foundation. The dominant question for production MAS becomes architectural — how to introduce protocol structure that does NOT rely on agents themselves recognizing convergence — rather than purely behavioral or training-based.