Control Agent Behaviors and Cost Beyond a Single Action: New Capabilities in Amazon Bedrock AgentCore

According to McKinsey, about 80% of organizations have encountered risky behavior from AI agents, making security and risk concerns the leading barrier to scaling agentic AI. To address this, AWS announced new capabilities in Amazon Bedrock AgentCore. The gateway, a fully managed serverless entry point for AI traffic, now supports temporal policies powered by Dogwood, an open source policy language built on Cedar. Temporal policies allow the policy engine to consider the agent's past actions in a session, enabling checks like blocking a transfer that uses a wrong account number or stopping purchases once a budget is reached. They are enforced deterministically at the gateway layer, denying by default and logging full context. Additionally, rate limiting is available, letting teams cap requests, tokens, and connection durations per user across tools, models, and agents, using identities from OAuth or IAM. Limits apply in per-second and per-minute windows and take effect without changing agent code. These capabilities aim to give enterprises control over agent behavior and cost, with policies and limits managed at the platform level.