PAGO: Prompt-Augmented Graph Organization for Multi-Agent Systems

Graph-based multi-agent language-model systems usually represent agents as nodes and communication channels as edges. This abstraction captures who communicates, but leaves the system's behavioral organization implicit: which operations are required, who should perform them, how they depend on one another, and how intermediate messages should be handled. We introduce PAGO (Prompt-Augmented Graph Organization), a framework that extends multi-agent organization from topology design to typed, executable graph composition. PAGO treats prompts as first-class graph primitives, jointly modeling agent and instruction nodes, their bindings and dependencies, communication edges, and message protocols. For each query, PAGO retrieves reusable primitives, selects a taskspecific graph under legality and resource constraints, and materializes it into executable agents, prompts, communication channels, and schedules while keeping the backbone language model frozen. Across seven benchmarks, PAGO with DeepSeek-v4-pro achieves an 81.58 seven-task macro-average versus 77.64 for the strongest non-PAGO baseline. Further analyses show task-dependent graph structures, robustness to candidate noise, and cross-backbone transfer, demonstrating that explicit prompt-level organization enables adaptive and executable multi-agent systems rather than optimizing agent connectivity alone.