Expansion-Contraction: A Multi-Agent Graph Traversal Pattern for Compound AI Systems

Expansion-Contraction is a domain-agnostic multi-agent coordination pattern in which an expansion phase dynamically spawns specialist agents mapped to nodes of a domain graph, and a contraction phase aggregates their findings inward toward a verdict. Agent topology emerges from data structure rather than hand-design, and each agent's context stays small regardless of graph size.

Abstract

Compound AI systems that coordinate multiple specialized agents offer a promising path for complex reasoning tasks, yet principled architectural patterns for multi-agent coordination over structured data remain under-explored. We introduce Expansion-Contraction, a multi-agent graph traversal pattern in which an expansion phase walks a domain graph outward from a query origin, dynamically spawning ephemeral specialist agents at each node, and a contraction phase aggregates their findings inward to produce a verdict. Agent topology emerges isomorphically from the data graph rather than being hand-designed, and each agent operates on a small local context—avoiding the context-window saturation that degrades single-agent approaches on large graphs. We instantiate the pattern for supply chain root cause analysis, integrating domain-specific tools with temporal lead-time propagation. Across eight datasets (three real-world, five synthetic with controlled depth and width), Expansion-Contraction achieves 98.2% accuracy on a production supply chain (624 cases) and 100% on public benchmarks, outperforming single-agent baselines by 14+ percentage points while degrading gracefully as graph complexity increases. A deterministic depth-priority disambiguation heuristic, motivated by our failure analysis, further improves Dataset A accuracy to 99.5% (621/624, 95% CI [98.6%, 99.9%]). To assess transfer, we evaluate the pattern on a second domain—microservice dependency tracing over a 17-service DAG (100 scenarios)—where Expansion-Contraction reaches 88% overall accuracy and 85% on NLP-complex cases (vs. 55% for the next-best baseline). Investigation caching reduces token usage by up to 93.9%, concurrent path analysis yields up to 1.43× speedup, and a production deployment demonstrates the pattern’s viability for enterprise-scale agentic systems.