Reading network state is the easy half. The real test of agentic operations is whether a model can safely change the network. In this session I pushed the lab from read-only intelligence into provisioning.
Intent in, configuration out
The workflow is simple to describe and powerful in practice:
- Natural-language input — “create four static routes from 1.1.1.0/24 on pe13 and pe23 in the INTERNET VRF.”
- Understanding — the model parses intent, identifies the devices, and resolves VRF context.
- Generation — it produces the exact Junos
setconfiguration. - Execution — the MCP server applies and commits the change.
- Verification — routes installed, BGP advertising, peers receiving.
What got built
In a single session, working through Claude and the MCP server, the lab saw:
- A Flex-Algorithm 128 deployment that would normally take 3–4 hours, done in about 15 minutes.
- Two L3VPNs provisioned across four devices — VRFs, 32 static routes, SRv6 micro-SIDs allocated automatically, and inter-POP connectivity verified.
- Network-wide BGP route analysis in seconds.
The honest takeaway
The point isn’t to remove the engineer — it’s to remove the toil. Memorizing syntax, hand-correlating data, and reactive troubleshooting are exactly the tasks that scale badly and burn people out. Hand them to an agent that works against validated tools and a real source of truth, and engineers get to spend their time on architecture and intent.
This is lab work, shared in the spirit of pushing the practice forward. If you’re experimenting with MCP and networks too, I’d love to compare notes — reach out.
Adapted and summarized from my original three-part LinkedIn series.