Your AI Agent Needs an Operating Environment. Who's Engineering It?
DevOps.com, Tuesday, November 3rd, 2026
Itential's John Capobianco argues AI agents doing infrastructure work need an engineered operating environment, not just a better model.
John Capobianco of Itential says the harness around an AI agent, meaning its tools, context, permissions, approvals and execution controls, is an architectural decision that someone must own.
Instructions in a prompt are not controls: mandatory steps such as approvals and verification must be enforced by the execution system.
Teams should define what context each task needs and how current it must be, and what the agent does when data is stale or contradictory.
Completion should require evidence such as verified state and recovered service, and the agent and its environment must be retested whenever models, tools or procedures change.