The first AWS guardrails I want around AI agents
AI agents change the tempo of cloud work. They make the old cloud basics more important, not less.
Before I let an agent workload get anywhere near production-shaped AWS access, I want a few boring controls in place.
A budget or alarm tied to the workload
Account-level budgets are useful, but they are often too broad to catch the specific experiment that got away from someone.
IAM that says what the agent is allowed to do
Agents should not inherit broad permissions just because the first experiment was easier that way.
Tags that make the bill explainable
If the bill arrives and nobody can tell which agent, owner, environment, or experiment caused the spend, the team has already lost useful time.
A stop condition for retries
Retry logic is where a small mistake can become a large bill.
A human review step
Agent-authored infrastructure should be treated as draft work. A human still needs to review the plan, permissions, blast radius, and cost controls.