The communication layer is the bottleneck
Five-way independent sampling burns five times the compute for one agent's worth of capability on long-horizon search. Training only a small communication hub via RL — leaving the agents themselves frozen — lifts per-agent BrowseComp accuracy from 36% to 58% E083. A more extreme version of the idea: wire two frozen copies of the same model through a 1%-parameter bridge between their hidden states and they invent a structured communication protocol (quiet on routine tokens, loud on semantically critical ones) from task loss alone, lifting arithmetic from 36% to 96% E040. Splitting a fixed parameter budget across three communicating agents nearly doubles the accuracy of one agent with the same budget E060, suggesting organisation itself is a scaling axis the field has been ignoring.
The Searcher/Navigator split for deep research is the same idea in different clothes: swarms of Searchers assemble a typed evidence graph, a single Navigator reads only the graph, and the 1200-to-1 compression ratio is what finally lets parallel scaling keep paying off E051.
Coordination and verification
Coordination bugs — not bad reasoning — are the dominant failure mode in multi-agent LLM systems, and standard testing almost never catches them. Wiring LLM protocol design into TLA+ model checking converges in four iterations or fewer across 48 tasks, and the interesting operational result is the capability buffer: verified protocols lose ~15 points of task completion under model downgrades while prompt-only approaches lose ~33 E034. Verification has shifted from correctness theater to a practical lever.
The regime change isn't the model checker — it's that LLMs can now cheaply draft the formal spec that used to be the bottleneck. A parallel result in distributed-systems verification compresses 9-12 months of expert work into ~10 hours of compute and sometimes produces verified implementations that run 3x faster than hand-written references, because joint code+proof synthesis pushes toward representations that are both verifiable and efficient E075.
Emergent collapse and paradoxes
Three LLMs talking for a thousand rounds grow vocabulary while their semantic content barely moves — about 3x more anchored than human Reddit threads — and twelve intervention categories (temperature, prompts, personas, model mixing, removing safety training, RL for diversity) all failed to break the pattern, with induction heads identified as the mechanistic cause E073. Counterintuitively, training models to be diverse made independent runs look *more* like each other.
The capability-vs-safety inversion in manager/worker setups is in the alignment topic E058, but its multi-agent flavour matters here: cooperation behaviour in self-play doesn't transfer — one weak model in a Prisoner's Dilemma group can unravel cooperation for everyone, a failure mode invisible to standard evaluation E018. And same-model-attacks-itself in multi-turn safety conversations reaches 100% essay production on against-consensus topics for ~$105 in API costs E045. The composite picture is that closed-loop multi-LLM systems have failure modes that aren't visible in any single-agent test.
Episodes anchoring this topic
- 083-agentfugue-agent-scaling-for-long-horizon-tasks-through-coll
Demonstrated that training the inter-agent communication layer beats scaling individual agents.
- 060-neuromas-multi-agent-systems-as-neural-networks-with-joint-r
Showed that splitting a fixed parameter budget across communicating agents nearly doubles accuracy.
- 034-tracefix-repairing-agent-coordination-protocols-with-tla-cou
Wired LLM protocol design into TLA+ model checking, showing verification as an operational lever.
- 040-the-bicameral-model-bidirectional-hidden-state-coupling-betw
Showed two frozen models learn to communicate through hidden states from task loss alone.
- 073-multi-llm-systems-exhibit-robust-semantic-collapse
Found robust semantic collapse in closed-loop LLM conversations resistant to twelve intervention categories.
- 075-inductive-deductive-synthesis-enabling-ai-to-generate-formal
Demonstrated that joint code+proof synthesis compresses expert verification work by orders of magnitude.