WaAgents: A Waterfall-Inspired Framework for Effective Multi-Agent Collaboration

ICLR 2026 Conference Submission25111 Authors

Keywords: Multi-Agent Systems, Large Language Models, Waterfall Model

TL;DR: We introduce WaAgents, a multi-agent collaboration framework inspired by the Waterfall model, which can improve the effectiveness of multi-agent systems in complex task resolution.

Abstract: Large Language Models (LLMs) have revolutionized the construction of multi-agent systems for complex problem solving, leveraging their prowess in natural language understanding for semantic parsing and intent recognition, alongside robust logical reasoning for intricate task execution. Despite these advances, prevailing LLM-based multi-agent frameworks suffer from a critical shortfall: the absence of explicit, predefined stage segmentation. This leads to pervasive information redundancy in inter-agent communications, manifesting as irrelevant discussions without focused topics, and exacerbates decision conflicts in free-discussion paradigms, where agents of equal status deadlock over divergent opinions, ultimately hindering effective resolutions. To address these limitations, we introduce WaAgents, a novel multi-agent collaboration framework inspired by the Waterfall Model in Software Engineering. WaAgents delineates the problem-solving process into four sequential, interdependent stages: Requirement Analysis, Design, Implementation, and Reflection. In the Requirement Analysis stage, Requirement Analysis Agents parse user intents to produce a structured task specification, facilitating downstream processing. Designer Agents in the Design stage then employ this specification to decompose the task into granular sub-tasks, systematically assigning them to dedicated Worker Agents. During Implementation, each Worker Agent executes its sub-task through targeted operations and computations. Anomalies trigger the Reflection stage, where Error Analysis Agents diagnose root causes, distinguishing design from implementation errors, and enact precise repairs, ensuring iterative refinement without disrupting workflow integrity. This stage-driven, highly structured workflow provides each agent role with explicit, concentrated objectives, which substantially mitigate information redundancy. Furthermore, by strictly enforcing the predefined flow, WaAgents fundamentally eliminates the decision conflicts inherent to free-discussion, thereby ensuring the coherence and effectiveness of the entire solution process. Empirical validation across challenging benchmarks, including mathematical reasoning and open-ended problem solving, confirms the efficacy and marked superiority of the WaAgents framework.

Primary Area: other topics in machine learning (i.e., none of the above)

Submission Number: 25111