Abstract
Long-horizon planning tasks present a "complexity ceiling" for Large Language Models (LLMs) due to context bloat and rigid coordination. This paper introduces a multi-agent system (MAS) that utilizes a structured shared notebook for information grounding and a reflective orchestrator for dynamic task routing. On the TravelPlanner benchmark, this architecture achieved a 25% pass rate, a significant improvement over the 7.5% baseline of single-agent systems.
Introduction
As LLMs are increasingly deployed as autonomous agents, they frequently encounter failure modes in complex planning scenarios. The authors identify two primary issues: Context Bloat, where excessive tool output leads to hallucinations of non-existent entities, and Rigid Workflows, where fixed pipelines prevent the necessary backtracking required to resolve interdependent constraints (e.g., adjusting a flight to accommodate a specific restaurant budget).
Methodology
The proposed MAS architecture moves away from messy chat histories toward structured collaboration through specialized components.
Key components include:
- The Notebook ($N$): A shared "source of truth" where experts record distilled facts, filtering out noisy raw tool outputs to prevent the "telephone game" effect.
- The Orchestrator ($D$): A dynamic agent that uses self-reflection to decide which expert to call next, allowing for non-linear planning and backtracking when constraints conflict.
- Expert Specialization: Specialized agents for Transportation, Hotels, and Dining handle raw data, passing only relevant information to the notebook.
- Plan-Refine Loop: A dedicated Plan Compiler generates the initial itinerary, which is then iteratively polished by a Plan Critic.
Experimental Results
The system was evaluated using the TravelPlanner benchmark, a sandbox containing roughly four million entries. The MAS, powered by Claude 4 Sonnet, reached a 25% pass rate, tripling the performance of single-agent baselines (7.5%).
Impact of the Shared Notebook
The implementation of the Notebook significantly mitigated the "hallucination cascade." By forcing agents to commit to structured facts, the system reduced entity-related errors by 18%.
As shown in Figure 2, the notebook was particularly effective in reducing errors in high-density data categories like Flights and Accommodations, eliminating more than half of the errors related to flight numbers.
Coordination and Backtracking
The Orchestrator's ability to backtrack was vital. On average, the system revisited previous steps 0.6 times per task to resolve conflicts. While this increased coordination complexity—leading to a slightly lower "Delivery Rate" (task completion within step limits)—the quality and constraint satisfaction of the resulting plans were substantially higher.
Conclusion and Future Work
This research demonstrates that structured memory and dynamic orchestration are essential for scaling LLM agents to long-horizon tasks. While the MAS is more computationally expensive, it successfully breaks the complexity ceiling of single-agent systems. Future work will focus on improving the efficiency of the Orchestrator to prevent coordination loops and developing more robust conflict detection mechanisms for deeply coupled constraints.
Summary updated