Autonomous System Restructures LLM Agents at Runtime

Traditional multi-agent LLM systems typically operate with a fixed team structure, which can lead to performance bottlenecks when individual agents become overloaded. This overload can stem from managing too many action categories, accumulating tool errors, or excessive queuing. To address this limitation, a novel runtime mechanism called Autonomous Topology Mutation (ATM) has been introduced, enabling multi-agent LLM frameworks to dynamically restructure themselves.ATM continuously monitors a "Bottleneck Index" comprising six key signals, including queue depth, context thrash, and tool-error rates. When this index breaches a calibrated threshold for a sustained period, ATM intelligently factorizes the overloaded agent into more specialized sub-agents. The original agent then transitions into a coordinator role, all while maintaining its external identity. This process is safeguarded by three critical invariants: capability monotonicity, ensuring no loss of function; state-routing completeness, controlling memory transfer based on privacy levels; and shadow-before-live validation, where new topologies are tested before deployment.Experimental results using DeepSeek-V3-driven tasks show a dramatic improvement in code-task success rates, from 3.3% to 61.7%, when the ATM factorizer split is applied. The full system also effectively reduces high-privacy memory exposure to zero events per task, without compromising task quality. Furthermore, the runtime rails for ATM's invariants add minimal latency, demonstrating its practical viability for building robust and adaptive multi-agent LLM systems.