A real AWS Strands Agents v1.54.0 agent — an Agent on an OpenAI-compatible model provider — runs in-process with the Mission SDK. No mock, no sidecar approximation.
AWS's authority model is IAM, whose defining rule is that an explicit Deny overrides any Allow. So we test the boundary immediately outside a Strands agent's tool calls, on the Wealth Manager workload. ReDevOps composes an organizational guardrail (SCP), the role's IAM policy and the Mission grant — and admits an action only when IAM allows it, no explicit Deny fires, and the Mission grants it. It runs in shadow first.
The real Strands agent proposes portfolio actions. ReDevOps denies each one an SCP guardrail explicitly denies, IAM doesn't allow, or the Mission doesn't grant — a permissive IAM Allow never overrides an explicit Deny — while admitting the 2 of 2 fully-permitted actions.
The Wealth Manager runs in shadow across 500 households: the agent proposes an action for each, ReDevOps holds all execution — zero real side effects — and measures what the policy would admit. Promotion to execution is a separate, governed step.
When the real Strands agent proposes withdrawing a client's cash to a personal account, the SCP guardrail denies it before any transfer — and the governed promotion fires an approved action exactly once, with 0 duplicates on replay.
The composed-policy evaluation is deterministic — real IAM/SCP semantics — and the value is the enforced mechanism. The real Strands agent genuinely proposes the actions, including one the guardrail then denies before anything happens.
// strands-agents 1.54.0 · Mission SDK 0.2.0 · Context Runtime 0.2.0 · local, no AWS cloud · github.com/redevops-io/redevops-integrations