ATM lifts multi-agent code-task success from 3.3% to 61.7% by splitting overloaded agents at runtime.

Sidik, Levi, and Kimhi argue fixed boot-time topology has no answer when one agent becomes the bottleneck, and propose a six-signal Bottleneck Index (queue depth, context thrash, tool-error rate, role entropy, retry-loop rate, cross-agent wait) that triggers factorizing an overloaded agent into sub-agents while hot-swapping the parent into a coordinator that keeps its external identity (arXiv 2607.20488). Across 720 DeepSeek-V3 runs, the factoriser split raises code-task success from 3.3% to 61.7% and privacy-aware state routing cuts detected high-privacy memory exposure from 2.0 to 0.0 events per task, at under 500µs p99 on the hot path.