When Cloud Agents Meet Device Agents: Lessons from Hybrid Multi-Agent Systems
Summary
This paper systematically studies hybrid multi-agent systems combining cloud-based LLMs and on-device SLMs, revealing task-dependent optimal architectures and challenging the assumption that more frontier compute always improves performance.
View Cached Full Text
Cached at: 05/29/26, 11:01 AM
Paper page - When Cloud Agents Meet Device Agents: Lessons from Hybrid Multi-Agent Systems
Source: https://huggingface.co/papers/2605.30102
Abstract
Hybrid multi-agent systems combining large and small language models offer flexible inference trade-offs, but optimal architecture depends heavily on specific tasks and performance metrics.
The design space of agentic AI inference spans two extremes: frontierlarge language models(LLMs), typically hosted in the cloud and offering strong performance across a wide range of tasks at substantially high cost, and more cost-efficientsmall language models(SLMs), which are amenable toon-device inference.Hybrid multi-agent systems(MASs) combining on-device and cloud models offer a promising middle ground, but they also introduce a complex and poorly understood design space in whichtask accuracy,monetary cost, andedge energy consumptionare tightly coupled; in the absence of general design principles, hybrid components, although not the most prevalent choice, are typically introduced through ad hoc decisions tailored to specific domains. In this work, we examine this design space more systematically. We adapt two representative MAS architectures to support hybrid inference and study how individual design choices shift the operating point along thePareto frontierof power, cost, and performance. Our findings paint a nuanced picture of hybrid MAS design: while SLMs can effectively benefit from LLM assistance, the optimal architecture is highly task-dependent, and greater frontier-level compute does not consistently translate to better performance.
View arXiv pageView PDFAdd to collection
Models citing this paper0
No model linking this paper
Cite arxiv.org/abs/2605.30102 in a model README.md to link it from this page.
Datasets citing this paper0
No dataset linking this paper
Cite arxiv.org/abs/2605.30102 in a dataset README.md to link it from this page.
Spaces citing this paper0
No Space linking this paper
Cite arxiv.org/abs/2605.30102 in a Space README.md to link it from this page.
Collections including this paper1
Similar Articles
Runtime-Adaptive Computational Mobility and Self-Deployment of LLM Agents
The article describes a personal experiment with LLM agents that can migrate between different hardware substrates based on task requirements, exploring concepts of agent identity and computational mobility.
Hybrid Reasoning Agents: LLM agents that think in latent space, act through text
The paper introduces HybridAgents, an LLM agent architecture that dynamically switches between latent-space reasoning and text-based action via a lightweight HybridLink, combined with HybridSFT and HybridRL post-training, achieving 10.4% performance gains with large token, latency, and cost reductions across 11 benchmarks.
More Agents, More Samples, or More Thinking? Untangling Coordination, Ensembling, and Test-Time Compute in Multi-Agent LLMs
This paper synthesizes 516 multi-agent LLM configurations to disentangle coordination gains from sample multiplicity and inference compute, finding that multi-agent systems match or underperform compute-matched baselines in 74% of comparisons, implying communication only helps with partitioned evidence, external verification, or model diversity.
Unified Agent: Managing Interactions across Devices
This paper introduces Unified Agent, a stateful AI agent that maintains compact interaction state across devices and time to handle cross-device, cross-time user requests, outperforming existing agent designs in a new benchmark.
Should SLMs be in the Driver's Seat? A Study of Local Models as Interpreters in Cloud-Written Agent Programs
论文研究了一种混合架构:云端大模型预先编写程序,将涉及用户私有数据的运行时判断委托给本地小模型(SLM)通过普通函数调用来解释。在 AppWorld 基准上,该方法在所有六种模型与数据分区组合中均提升了纯代码 agent 的成功率(最高 14.9 个百分点),在保持数据本地化的同时缩小了与云端 ReAct 循环的差距。