AWS outlines four layers of control for autonomous AI agents

Scope and enforceable limits

In an October 15 post, AWS set out its Trust in Depth approach: define an agent’s job, identify who authorizes it, enforce boundaries during execution and examine what happened afterward. The post recommends expanding autonomy as teams establish evidence that the relevant rules are enforced and failures can be detected and handled.

For teams building with AgentCore, the Policy documentation describes checks outside the agent’s own code. Requests passing through AgentCore Gateway are evaluated before tools can be accessed. Rules can depend on identity, tool arguments and earlier actions in a session, for example requiring approval before a transfer. This gives operators a concrete place to enforce workflow prerequisites.

Identity and task performance

AgentCore’s Identity documentation describes workload identities and the authentication, authorization and credential management needed when agents access AWS resources or third-party services on a user’s behalf. For a deployment review, this makes the agent’s identity and its route to service credentials explicit questions to resolve alongside the application’s intended task.

Finally, the Evaluations documentation explains that agent traces are scored using language models as judges, with built-in or custom evaluators. The service supports checking task performance, edge cases and consistency before and after deployment. A useful question for buyers is which tasks and failure cases are actually evaluated. These are model-based assessments, so their scores need interpretation in the context of the work the agent is expected to perform.