AWS launches Trust in Depth: verifiable layered controls give AI agents more autonomy

AWS's Trust in Depth is a layered governance framework covering four stages of an agent's lifecycle: Scope before running (defining the agent's responsibilities, accessible tools and data, and mandatory rules); Know at deployment (assigning each agent an identity, tracing authorization origins, and building a managed catalog via AWS Agent Registry); Bound at runtime (enforcing policy outside the agent's reasoning through AgentCore Gateway, with time-based policies that dynamically adjust permissions as a session progresses); and Observe after running (Amazon CloudWatch Omni capturing every step's trace, with built-in evaluators scoring correctness and faithfulness). The core mechanism is automated reasoning: for precisely expressible rules like 'only managers can approve access,' the system can prove them true across all possible paths or pinpoint failure cases; for judgment-based rules like 'keep a friendly tone,' it restricts agent permissions and hands off to humans. The technology has run for over a decade in systems including IAM (over 2.1 billion requests per second), the Nitro Hypervisor, and S3 Block Public Access. For developers building agents with Strands or Kiro, the Strands Box sandbox enforces boundaries outside the agent's instructions, while Kiro uses automated reasoning during coding to find requirement contradictions and generate property tests. For enterprise ops teams managing multi-cloud agents, these mechanisms mean agent autonomy can be progressively expanded without sacrificing security.