OpsMill Intent: AI Ops You Can Trust

Xavier Martins-Rivas, field CTO of OpsMill, explained how Infrahub’s intent model provides a trusted foundation for AI operations. The session covered moving institutional knowledge out of individual engineers and into structured, versioned, validated data, a top-down knowledge graph linking services to configurations and devices, and guardrails that let AI agents act on infrastructure safely. Martins-Rivas presented the segment at Networking Field Day 41, noting that Infrahub now has more than 600 live deployments.

Martins-Rivas argued that most tools reflect their builders’ assumptions, so organizations fill the gaps with “Bob,” the long-tenured engineer who knows why a given switch must never be touched. That works until Bob goes on vacation. Intent replaces that tribal knowledge with structured data, shifting from teams carrying complexity to teams managing trust. He contrasted observability-style knowledge graphs, which build upward from devices and struggle to map impact onto custom business services, with Infrahub’s approach. Infrahub starts at the service layer and works down to configurations and devices, so users can trace relationships in both directions. He described four pillars of trusted infrastructure: a schema that models everything, versioning to move forward and back, validation against the organization’s own requirements, and generators and transforms that produce configurations and artifacts. When a delegate asked about test frameworks, he said validation rules are entirely user-defined to match each use case.

For agents, Martins-Rivas described two modes. Existing skills let agents build new services, such as an L3VPN, while MCP servers answer questions like what depends on a given device. To guard against hallucination, every agent change goes through a branch and a validation process that can plug into existing workflows like Jira tickets. He compared this to laying train tracks with gates so agents can’t go off the rails. Critically, Infrahub, not the agent, holds the memory. Its immutable database records every change and who made it, enabling rollback, and agents can be swapped or retired without losing anything.