StigmergyRouter: A Fault-Aware Adaptive Routing Demo for Multi-Agent AI Systems
Jing Du (Northeastern University), Hang Zhao (Northeastern University), Kenneth Huang (University of Pennsylvania)
System Optimization & Efficiency Architectural Patterns & Composition
Summary
A pheromone-memory-based routing layer that adapts multi-agent routing from low-cost heartbeat feedback alone, improving failed-agent avoidance to 95.7% under specialist faults.
Description
As compound AI systems increasingly rely on multi-agent architectures, routing user queries to specialized agents becomes a critical bottleneck. Current approaches, from static keyword rules to LLM-call-based and heavier classifier-style meta-routers, either lack adaptability or add extra inference overhead before downstream work begins. We present StigmergyRouter, a fault-aware routing layer that maintains clustered pheromone memory over semantic embeddings and updates its routing preferences from low-cost execution heartbeat feedback alone, without labeled routing data. We evaluate across HotpotQA, GSM8K, and CNN/DailyMail with 150 queries per trial over 10 random seeds. We do not present StigmergyRouter as the strongest general-purpose router. The contribution is narrower: a low-supervision, heartbeat-driven adaptation layer that can sit on top of strong semantic routing and improve behavior under specialist failure. We therefore evaluate both the pure mechanism and a practical HybridStigmergyRouter that combines semantic priors with heartbeat-driven pheromone memory. Relative to a strong Semantic+Failover baseline, the hybrid variant gives up a small amount of routing accuracy (89.5% vs. 92.7%) while improving downstream task score (0.387 vs. 0.369) and failed-agent avoidance under specialist faults (95.7% vs. 83.5% for math faults). We show these behaviors in an interactive Streamlit demo with live fault injection and inspectable pheromone state.
