Control agent behaviors and cost beyond a single action: new capabilities in Amazon Bedrock AgentCore

This article announces new capabilities in Amazon Bedrock AgentCore to address trust and security concerns in autonomous AI agents through infrastructure-level controls.

• Temporal policies evaluate sequences of agent actions, not just individual requests, using Dogwood policy language
• Policies can enforce prerequisites, budget limits, action ordering, and human approval requirements at the gateway layer
• Rate limiting on the gateway caps token consumption, request volume, and connection duration per user across all tools and models
• Controls are enforced outside agent code, preventing agents from reasoning around or circumventing policies
• Dogwood is open source under Apache 2.0, providing visibility into policy evaluation

These infrastructure-level controls enable enterprises to scale autonomous agents with consistent security and cost governance without rebuilding controls for each agent.