Title: Adaptive Strategy Selection for Coordination in Enterprise Multi-Agent Frameworks

Abstract:

As enterprise multi-agent systems become more prevalent, they offer a variety of coordination patterns. However, operational deployments frequently lack empirical guidance on when to implement specific methods such as consensus, debate, or synthesis, versus opting for a simpler single-agent workflow. This study investigates whether coordination strategies ought to be chosen dynamically based on the specific problem class, rather than applying a uniform global setting.

To test this, we employed a controlled matrix comprising 30 enterprise tasks distributed across six industries and five problem classes. The experimental design included four distinct execution conditions, with three replications per cell and four model arms: qwen_local, sonnet, gemma_openrouter, and an auxiliary cloud-validation arm using OpenAI. A total of 1,440 outputs were generated and evaluated against a fixed Sonnet-based rubric.

The primary results are limited in scope but offer practical operational value, though they do not fully substantiate the original strict hypothesis (H1). The pre-registered criterion regarding exact winners and confidence intervals was not supported; the identity of the winning strategy proved unstable across different model arms, and several predicted approaches were merely close to, rather than definitively superior to, the best-performing alternatives. However, a weaker hypothesis suggesting that the predicted strategy remains "near-best" is strongly validated. Across every pre-registered model arm and problem class—and confirmed by the auxiliary OpenAI validation arm—the predicted strategy fell within 0.10 quality-score points of the optimal observed condition.

An notable exception to the original mapping involves structured compliance verification, where all model arms preferred the single_agent approach over consensus. Additionally, a pre-registered Kendall's W test revealed no significant difference in how consistently the four coordination conditions were ranked between Vietnamese-domain and English-domain tasks (mean W = 0.20 in both strata; signed-rank p = .85), leading to the rejection of H2.

We conclude that enterprise coordination policies should adopt dynamic routing as a calibrated default mechanism, rather than treating it as a deterministic rule for selecting a single winner.

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