Title: Maximizing Returns on Parallelism: Cohesion-Driven Task Partitioning for Multi-Agent Coding

Abstract: Multi-agent Large Language Model (LLM) frameworks enable the breakdown of intricate processes, such as software development, by leveraging parallel execution and isolating context. Yet, the practical deployment of additional agents brings communication overhead that can erode efficiency gains and increase expenses. This study reframes multi-agent orchestration as a graph partitioning challenge, balancing the ratio of communication to computation. While decomposing tasks can reduce the length of the critical computational path, dependencies between agents necessitate expensive context transfers. We apply this framework to repository-level software engineering through the introduction of Cohesion-aware Coder (Co-Coder). This system constructs dependency graphs using static analysis, isolates structural hub files, partitions the graph via community detection, and manages execution through a scheduler sensitive to dependencies. Evaluated across 28 real-world tasks on DevEval and CodeProjectEval, Co-Coder improves upon the Pareto frontier established by sequential methods, file-based parallel baselines, and Claude Code with Agent Teams. It achieves a pass rate increase of up to 14.0%, a wall-clock speedup of up to 2.10x, and a cost reduction of up to 35% via API usage. The most significant improvements are observed in projects with high dependency density. Co-Coder illustrates that cohesion-aware orchestration can render parallel coding agents both theoretically sound and practically efficient, offering a generalizable design principle for multi-agent systems.

Source: arXiv Generated at: 2026-06-02 00:00:00 UTC