Collaborative AI Agents and Critics for Fault Detection and Cause Analysis in Network Telemetry

Researchers Syed Eqbal Alam and Zhan Shu have introduced a novel framework for collaborative AI in a paper titled 'Collaborative AI Agents and Critics for Fault Detection and Cause Analysis in Network Telemetry.' The core innovation is a federated multi-agent system where specialized AI 'agents' perform tasks—like detecting a network fault—and separate AI 'critics' evaluate the agents' work. All communication is routed through a central server, with no direct agent-to-agent or critic-to-critic chatter. This design keeps each component's internal cost functions or their derivatives private, a crucial feature for security and competitive environments.

The system is designed for multimodal tasks, meaning it can handle diverse data types such as network telemetry, medical images, and text. The authors use multi-time scale stochastic approximation techniques to mathematically guarantee the system's convergence, ensuring the collaborative process stabilizes over time. A significant efficiency claim is that the communication overhead scales only with the number of data modalities (O(m)), not with the number of AI agents or critics, making it highly scalable.

In their evaluation, the researchers demonstrate the framework's efficacy specifically for the complex problem of network telemetry analysis. Here, it can autonomously detect faults, assess their severity, and diagnose the root cause. The proposed architecture has broad implications, extending beyond networking to applications like text-to-image generation, video synthesis, and healthcare diagnostics, where multiple AI models must work together reliably and privately.