Autonomous Topology Mutation Lifts Multi-Agent Code-Task Success From 3.3% to 61.7% by Splitting Overloaded Agents at Runtime

arXiv 2607.20488·medium signal

Sidik, Levi, and Kimhi argue that multi-agent LLM frameworks fixing team topology at boot time have no answer when one agent becomes overloaded, and propose ATM: a six-signal Bottleneck Index (queue depth, context thrash, tool-error rate, role entropy, retry-loop rate, cross-agent wait) that triggers factorizing the overloaded agent into sub-agents while hot-swapping the parent into a coordinator that keeps its external identity. Across 720 DeepSeek-V3-driven runs over four ablation conditions and three workloads, the factoriser split raises code-task success from 3.3% to 61.7%, and privacy-level-aware state routing cuts detected high-privacy memory exposure from 2.0 to 0.0 events per task. Three safety invariants gate every mutation — capability monotonicity, state-routing completeness, and shadow-before-live validation — at under 500 microseconds of p99 latency on the hot path, with implementation and traces open-sourced.