Aggregation-Guided Buffering and Orchestration for Wide-Area Mixture-of-Agents in the Optical Internet of Agents
Year: 2026
Type: article
Abstract: Wide-area Mixture-of-Agents (MoA) collaboration requires the exchange of large payloads over the Optical Internet of Agents (O-IoA) under fluctuating traffic conditions. To increase agent participation within collaboration deadlines under network resource contention, we propose the Aggregation-Guided Buffering and Orchestration (AGBO) framework. AGBO coordinates hierarchical update aggregation, intermediate buffering, and optical transmission using a capacitated time-expanded resource graph (TERG). We develop analytical models showing that temporal buffering reduces collaboration failure probability at the cost of higher candidate-path search complexity. We also develop an AGBO algorithm combining two-stage maximum-flow orchestration with Meta Agent (MA) selection. Simulation results show that AGBO retains more Work Agents (WAs) across different wavelength-utilization levels, payload sizes, and collaboration deadlines. At a wavelength utilization of 60%, it improves the WA participation ratio by at least 57 percentage points over the two baselines. In a convolutional neural network (CNN)-based collaborative-learning proxy, AGBO also achieves faster accuracy improvement and higher global test accuracy alongside higher WA participation.
DOI: 10.64509/jicn.24.140
Institution: Fuzhou University
Language: English
Subfield: Electrical and Electronic Engineering
Field: Engineering
Domain: Physical Sciences
Open Access status: hybrid
OpenAlex ID: W7225982873
