Amazon Bedrock AgentCore Introduces Advanced Control Capabilities for AI Agents

Amazon Bedrock AgentCore has introduced significant new capabilities aimed at improving the control, security, and cost management of AI agents at scale. These features include the implementation of temporal policies, powered by Dogwood, a new open source policy language specifically designed for AI agents, and rate limiting on the gateway.

Temporal policies extend existing controls by allowing teams to enforce rules over sequences of actions, rather than just individual requests. This means that agents can be restricted from performing certain actions based on their previous steps, such as preventing a transfer to an incorrect account or stopping a purchase once a budget limit is reached. These policies are enforced outside the agent’s own code, at the gateway layer, ensuring consistent application of security boundaries.

Dogwood, the new policy language powering these features, embeds Cedar and introduces temporal constructs for agent governance, including rate limits, time windows, prerequisite steps, and escalation triggers. It is available as an open source specification and reference implementation under Apache 2.0, providing transparency and supporting ecosystem development.

In addition, rate limiting capabilities allow organizations to set ceilings on token consumption and request volume per user, across all tools, models, and agents managed through the gateway. This helps control costs and prevent unintended resource exhaustion, regardless of how agents behave or how many retries they perform.

These enhancements aim to address the growing need for trust and security in autonomous AI systems, especially as models become more capable and take more consequential actions. By moving control logic into the platform infrastructure, Amazon Bedrock AgentCore enables enterprises to deploy more autonomous agents with confidence, without requiring rearchitecting existing systems.

The company plans to continue investing in identity, observability, evaluation, and traceability features, further strengthening the system’s ability to bound agent actions and resource consumption. These capabilities are available now and can be adopted independently, offering a scalable way to enhance enterprise AI deployments.