AWS unveils Trust in Depth to let AI agents act alone

AWS frames the problem plainly: the same autonomy that makes agents useful also makes them unpredictable, since an agent picks its own tools and steps and may take a different route each run. The standard fix — a person approving every action — breaks down as approvals pile up and reviewers stop catching what matters, leaving capable agents idle.

The company positions Trust in Depth as a shift in purpose from the industry's long-standing Defense in Depth discipline. Where that approach limited what could go wrong, this expands what can safely be allowed. AWS says it has run automated reasoning in production for more than a decade, including IAM at more than 2.1 billion requests per second, and is now bringing it to agents. It draws a line between rules that can be proven, such as who may approve access, and those that cannot, such as keeping a friendly tone, which stay with people.

A worked example describes a help desk agent offboarding an employee named Sam. The agent blocks his sign-in, converts his mailbox, places a legal hold, and changes his license, while Policy in AgentCore stops steps taken out of order. AWS says every step is recorded so the team can verify the outcome.