Abstract:
Objective[Objective] We report production-running evidence for a star-topology system in which a single human operator routes six large language model (LLM) agents, examining whether an extreme design with no inter-agent communication channels can sustain everyday knowledge work.
Methods[Methods] Six mutually independent agents span two commercial platforms; cross-agent coordination is realized through rolling state documents—each workflow maintains a single authoritative Markdown document, append-only within a session and wholly replaced across sessions, distributed manually by the operator; memory is organized into four tiers, with the operator serving simultaneously as router, arbiter, and verification gate.
Results[Results] We formalize five rules governing document flow; handoff cost drops from being proportional to session length to proportional to increments; eight recurring failure modes and their operational rules are distilled from production incidents; cost analysis shows that for individual workers, push-based human routing outperforms autonomous orchestration frameworks in reliability-to-complexity ratio.
Limitations[Limitations] Evidence comes from a single operator in a single domain, is self-reported, and lacks third-party measurement or controlled comparison; the architecture has not yet been replicated by a second operator, and whether routing bandwidth can scale by orders of magnitude remains an open question.
Conclusions[Conclusions] For individual workers with asynchronous task flows and limited verification bandwidth, human-routed coordination with rolling state documents as the bus is a feasible architecture that attains reliable cross-agent coordination at minimal infrastructure cost.