Would you put AI agents in a separate AWS account?

A separate account can make the boundary much easier to see and review. Keep the agent’s working account narrow, then grant access to other accounts only through explicit cross-account roles with limited permissions. That makes sensitive access and destructive actions easier to inventory, though it does not make them safe by itself. The roles still need careful scoping and monitoring.

I’d place the agent alongside the service it belongs to unless it has a genuinely different risk profile. Separate accounts add operational overhead, and agents still need permissions to do their jobs. A well-scoped role, isolated runtime, and clear limits may be enough for a low-risk internal feature.

There are practical reasons beyond security. A dedicated account can make costs, quotas, and usage easier to track, and keep agent activity from getting lost among production workloads. That may matter if one workload can use up shared capacity or if you need a clean view of what the agent is doing.

I’d base it on the data and actions involved. If agents handle sensitive information, have broad tools, or fall under stricter isolation requirements, separate accounts are a sensible default. If they only operate within one application’s existing boundaries, keeping them there may be simpler. Either way, define the permissions and cross-account access deliberately.