Empowering Edge Agents: A Systematic Analysis of Edge-Cloud Collaboration in LLM Agent Execution

Empowering Edge Agents. A Systematic Analysis of Edge-Cloud Collaboration in LLM Agent Execution. By Pengcheng Huang and seven others, from Northeastern University in Shenyang, Tsinghua University, and one other institution. Published on the thirteenth of October. Its code repository is public and real: about thirty source files.

Edge-cloud collaboration tries to get the best of two places. Running on the edge brings lower inference overhead and keeps data local. Cloud-hosted models bring stronger capabilities. For a single query, deciding whether to call the cloud is a one-shot choice. Language model agents, though, increasingly work in long-horizon scenarios, and there the cloud's participation is interleaved throughout the run. That interleaving has effects at the level of the whole trajectory, and those effects remain insufficiently understood.

EC2-Bench is built to measure them. It evaluates edge-cloud collaboration in long-horizon agent execution along three axes at once: task performance, resource usage, and privacy exposure. Holding all three in one framework matters, because each style of collaboration can buy one at the expense of another.

The benchmark provides executable agent tasks set in a digital workspace, and a unified evaluation framework covering those three axes. With it, the authors empirically study diverse forms of edge-cloud collaboration, set against edge-only and cloud-only execution. The forms differ in who holds the wheel. In some, cloud participation is controlled externally. In some, execution is handed directly to the cloud. In others, the edge model itself controls when and how the cloud is brought in. Because the tasks are long, each of these choices is made again and again within a single run, and the benchmark scores the whole trajectory rather than any one call.

The study reports three findings. First, collaboration strikes a favorable balance among performance, resource usage and privacy exposure, compared with running only on the edge or only in the cloud. Second, as edge models get stronger, collaborative execution approaches cloud-only performance, and the differences between strategies shrink. Privacy exposure does not follow that convergence: it remains strongly dependent on strategy.

Third, stronger edge capability changes which collaboration is preferred. The preference moves away from cloud participation controlled from outside, and from handing execution straight to the cloud, toward collaboration the edge controls itself. Further analyses identify two edge-side abilities as important for collaborating well: deciding when to invoke the cloud, and following instructions.

The second finding carries the practical warning. As edge models improve, strategies start to look alike on task performance, so performance alone stops being a good way to choose between them. Privacy exposure keeps them apart, and so does the question of who should be in control, which the third finding says depends on how capable the edge model is. That is why measuring all three axes together, over whole trajectories, is the core of what the authors offer for designing future edge agents.