🚨 BREAKING: researchers planted a single bad actor inside a group of LLM agents. the whole network failed to reach consensus. this is the Byzantine Generals Problem. a 40-year-old distributed systems nightmare. and it's now your agent pipeline's problem too. in fully benign settings, with zero bad actors, LLM agents still fail to converge on shared values. and it gets worse as you add more agents to the group. the failure mode is revealing. it's not subtle value corruption. it's not one agent sneaking in a wrong answer. the models just... stall. they time out. they go in circles. the conversation never lands on agreement. this matters because the entire multi-agent AI hype assumes coordination works. autonomous agent swarms, collaborative problem-solving, decentralized AI systems. all of it assumes that if you put multiple LLMs in a room and give them a protocol, they'll converge on a shared decision. Byzantine consensus is one of the oldest, most studied problems in distributed systems. classical algorithms solved it decades ago with strict mathematical guarantees. the question was whether LLM agents could achieve the same thing through natural language communication instead of formal protocols. the answer, at least for now, is no. and the reason is worth sitting with. traditional consensus algorithms work because every node follows an identical deterministic protocol. LLMs are stochastic. the same prompt produces different outputs across runs. an agreement that holds in round 3 can dissolve in round 4 as agents revise their reasoning after seeing peer responses. this is the fundamental mismatch: consensus protocols assume deterministic state machines. LLMs are the opposite of that. it also means that "more agents = better answers" has a ceiling nobody's measuring. at some group size, coordination overhead and convergence failures outweigh any benefit from diverse perspectives. the practical implication is uncomfortable for anyone building multi-agent systems for high-stakes tasks. reliable agreement isn't an emergent property of putting smart agents in conversation. it has to be engineered explicitly, with formal guarantees, not hoped into existence. we're deploying multi-agent systems into finance, healthcare, autonomous infrastructure. and the consensus problem, the most basic coordination primitive, isn't solved yet.