Can Edge LLMs Team Up? A Fine-Grained Diagnosis of Multi-Agent Tool Calling

Whether edge-scale language models can call tools, and whether teams help, is understudied. We present a fine-grained diagnostic evaluation of tool calling on open-source edge models (0.6B–9B). The single-agent framework measures four dimensions and attributes every run to one of seven failure modes; parameter filling, not tool selection, separates single-agent models. The evaluation extends to multi-agent systems: an orchestrator-workers workflow over 30 MCP tools, a two-layer failure taxonomy of role delegation and worker execution, and a noise-floor protocol of 30 repeats with paired t-tests, exact permutation tests, and bootstrap intervals. Delegation rises steeply with orchestrator scale. Even a Qwen3-0.6B coordinator can delegate: a template-controlled sweep traces its zero delegation to the chat template attached to the model tag rather than to weight precision, and the same weights reach 12.5–22.1% under the template embedded in the model file. Orchestrator-workers teams do not significantly outperform a matched single model or best-of-3, an oracle-aided upper bound rather than a deployable selector, at any scale we tested: delegation overhead consumes the accuracy workers achieve. These findings hold on 4090 GPUs and reproduce on an Ascend 310P NPU. Cross-family runs reproduce the general directional pattern, and the tested orchestrator-workers teams provide no demonstrated advantage over the corresponding single-model baselines under this benchmark and deployment setup.