Giving an AI agent access to production infrastructure is easy to demonstrate. Making sure it doesn't turn a small issue into a major outage is a much harder engineering problem.

Imagine an agent detects rising storage latency and decides to fix it. It identifies a possible bottleneck, selects a remediation action, and executes it. But what happens if the diagnosis is incomplete, the system is already under pressure, or the proposed fix makes things worse?

In traditional infrastructure operations, we have learned to be careful with automation. We use health checks, change windows, canary deployments, access controls, rollback procedures, and human approvals. Agentic AI doesn't eliminate the need for these safeguards. It makes them even more important.

In this session, I'll explore how established infrastructure engineering practices can help us build safer agent-driven operations. Using a hypothetical storage incident and a reference architecture based on Python, Ansible, monitoring systems, and policy-controlled tool access, I'll walk through how an agent could move from observation to recommendation and eventually to a controlled action.

We'll look at how to:

- Separate what an agent can observe from what it is allowed to change.
- Use health checks, policies, and approval gates before taking action.
- Prevent repeated or conflicting remediation attempts.
- Recognize when automation should stop, escalate, or require a human decision.
- Design actions that are observable, auditable, and reversible wherever possible.

This isn't about replacing infrastructure engineers with AI agents. It's about applying the lessons we've learned from running complex production systems to a new generation of automation.

Attendees will leave with a practical design pattern they can adapt when connecting AI agents to enterprise tools, infrastructure APIs, and operational workflows.