A new research demo shows AI agents can manage optical network operations from end to end autonomously, using the same tool-calling standard that underpins most modern AI assistants.
Researchers, whose names and institutional affiliation are not included in the source abstract, built an MCP-enabled agentic AI system to handle IPoDWDM (IP over Dense Wavelength Division Multiplexing) networks without tying operators to any single vendor's stack. The system pairs GNPy, the open-source optical network planning tool, with live telemetry to execute multi-layer lifecycle automation. Closed-loop control lets the agent detect a network condition and respond without waiting for a human to authorize each step. Validation ran on a real testbed, not a simulation.
MCP, the Model Context Protocol introduced by Anthropic as a standard for connecting AI models to external tools, has mostly appeared in developer productivity and enterprise chat contexts so far. Extending it to carrier-grade optical infrastructure is a significant jump: IPoDWDM networks carry backbone internet traffic, and autonomous decisions at that layer carry real consequences. The vendor-agnostic framing is also notable; telecom procurement is notoriously siloed, and an automation layer that works across vendors has obvious commercial appeal.
A testbed demo is more credible than pure simulation, but the gap between a controlled lab and a live production network has quietly buried plenty of promising automation projects before.