LLM‑driven agents are being employed to accomplish complex long‑horizon tasks without human supervision. To advance toward the task goal, agents must take on duties such as planning and verification during execution. We first study how a single agent distributes these duties over a long horizon and find that verification correlates most strongly with success, yet it typically occurs late, leaving little room for correction. At the same time, as records from different duties accumulate in a shared context, the proportion of duty‑relevant information diminishes. This motivates us to ask whether a multi‑agent system can alleviate the issue by assigning separate contexts to different duties. Experiments show that context separation alone is insufficient: agents may lose awareness of the overall goal, repeat previous investigations, and receive irrelevant information without the necessary background. To address these shortcomings we propose Goal‑Aware Dynamic Multi‑Agent Collaboration (GAD). GAD organizes subtasks and their dependencies in a growing execution graph. Each context retains the original instruction, the assigned task’s goal, and its acceptance conditions, keeping the agent focused. During execution, agents can split tasks into subtasks and declare dependencies based on observed evidence; these dependencies dictate when agents work and which results they receive. Verification and recovery modules check subtask outcomes and overall task completion against their goals, providing feedback for repair when checks fail.
Review