CASA: Capability-Aware Agent Discovery and Binding for Network Operating Systems

AI-driven network operations increasingly use agents for tasks such as BGP analysis, telemetry analysis, topology inference, and incident diagnosis. However, today's network operations are often tied to the specific agents performing these tasks; so each operation must handle agent failures or changing conditions itself. We present CASA, a discovery and replacement layer for agentic network operating systems. With CASA, a network operation specifies what task it needs performed and any operational requirements, rather than naming a particular agent. CASA finds a suitable agent, selects one that currently satisfies those requirements, and replaces it if it becomes unavailable or no longer meets them. Across registries containing 10 to 1,000 agents, median discovery latency increases from 0.75 ms to 77.37 ms, with a 95th-percentile latency of 82.77 ms at 1,000 agents. At that scale, median agent selection and task assignment take 2.33 ms and 0.03 ms, respectively.