On 6 August 2026, AWS described two governance capabilities for Amazon Bedrock AgentCore: temporal policies and rate limiting. Temporal policies can authorize a later action using earlier events in the same session. Rate limits can constrain requests, inference tokens and concurrent connections. The change is about a gateway control plane for multi-step agents, not a new model.
Policies that read the session
A temporal rule can express ordering, require a later argument to match an earlier result, require recorded human approval or check data freshness, according to AWS. The policy sits outside the agent code and evaluates the session history before a later action proceeds. That gives teams a concrete way to test whether a workflow can move from action A to action B only under the intended conditions.
Budgets at the gateway
The same material describes limits scoped through OAuth or IAM and applied to requests, inference tokens and concurrent connections, with per-second or per-minute windows. AWS says these controls do not require changes to agent code. In practice, that could give an operator a place to bound retries and runaway usage while keeping the agent logic separate from the policy layer.
Claims to test before production
AWS presents the controls as deterministic, deny-by-default and logged. Those are vendor claims. The supplied evidence does not independently establish enforcement under production load, latency or gateway overhead, false-positive behavior, availability, pricing, regional coverage, published limits or total cost. The linked Dogwood project is a policy language and reference implementation, and its README says it is not for production.
A rate limit can bound requests or tokens; it does not show that an agent selects good actions. A recorded approval is not proof that identity, data flow and surrounding permissions are sound.
A small policy harness
- Use a disposable gateway or the Dogwood examples with synthetic data.
- Define a two-step flow where B must use a value produced by A and a human approval is required.
- Run the valid sequence, then try B before A, alter the earlier result and omit approval.
- Check the denial reason, audit record and latency for each attempt.
- Set a deliberately low rate limit and run a controlled retry loop to identify which boundary triggers.
Watch for production semantics, region and pricing details, service limits, and independent measurements of enforcement overhead. The useful follow-up is a reproducible report showing which violations were blocked and what the controls cost under realistic retries.
Sources: AWS Machine Learning Blog, AWS temporal policies announcement, Dogwood and AgentCore developer guide.