Proxifield Routes Multi Agent Teams by Semantic Proximity at Scale
An arXiv paper shows decentralized routing beats centralized star topologies as agent teams grow, retaining 73.6% reward under severe permanent failures.
What changed
Researchers posted Proxifield on arXiv on 16 September 2026 (2609.20889). The protocol builds sparse communication graphs at inference time from semantic proximity among large language model agents, without centralized planners or extra training. Four routing signals drive links: direct address, information needs, plan alignment, and information complementarity.
On drone search and rescue and the HiddenBench collective reasoning benchmark, Proxifield beat centralized star routing and a decentralized shared context baseline. As team size rose from five to fifty agents, its task reward advantage over star widened from 5.4% to 59.5%. Under the harshest permanent failure scenario tested, Proxifield retained 73.6% of its no failure reward versus 38.8% for star and 58.3% for shared context.
Why it matters
Most enterprise agent stacks still mirror star topologies with a single orchestrator. That creates a coordination bottleneck and a single point of failure when agents drop offline. Proxifield offers a training free alternative that scales with team size, which matters for warehouse robotics swarms, incident response copilots, and multi site research automation.
Performance improved with model scale in the paper’s ablations, from a 35B parameter base to a 397B parameter model, suggesting the routing signals benefit from stronger encoders.
Who is affected
Platform engineers building multi agent workflows; reliability owners running agent fleets in regulated environments; researchers comparing coordination protocols before committing to reinforcement learning heavy orchestration layers.
What to do next
Pilot a decentralized routing layer on one high fanout workflow and measure fault tolerance when you simulate permanent agent loss, before expanding star shaped orchestrators.
What to watch
Follow up benchmarks on real enterprise tool use traces and whether vendors adopt semantic proximity routing in shipping agent frameworks.
Sources
- Primary. Sannia et al., Proxifield: Decentralized Multi Agent Communication through Semantic Proximity (16 September 2026). Establishes routing design, benchmarks, and failure robustness numbers.