SwiNOG

Swiss Network Operators Group

From Lab to Production: The Hard Parts of Agentic NetOps

At SwiNOG#41 we explored what agentic AI can and can’t do in network operations. The talk sparked many follow-up conversations. Most of them circled the same question: what happens once you leave the lab? Based on NetFabric deployments in Swiss ISP and MSP networks, this talk walks through the main roadblocks we hit moving agentic NetOps from small-scale tests to production networks, and what it took to get past each of them. We focus on three areas:
– **Correctness:** In production, misunderstanding an interface state or firewall rule is only one of many ways an agent can fail. Issues like incomplete investigations or overconfident answers matter just as much. We highlight failure cases drawn from real deployments and show how we addressed the most pressing ones.
– **Security:** How can an agent safely consume the output of CLI commands on production devices? And what can go wrong when it has direct access to a production database?
– **Model choice:** New models and agent paradigms appear every month. We show why simple tests produce misleading results and what a robust evaluation framework looks like.