AWS open-sources Dogwood, a policy language that governs AI agents based on their history of actions

AWS released Dogwood, an open-source governance language for Amazon Bedrock AgentCore that extends Cedar-based, point-in-time authorization with temporal conditions. Where existing policy tools evaluate only a single request in isolation, Dogwood looks back over an agent’s recent tool calls to enforce rules like required approvals, rate limits, spend caps, and orderings – the exact sequence-level constraints that could catch runaway or coordinated agent behavior. It’s built on formal temporal logic, is Cedar-compatible with no migration needed, and is released under Apache 2.0.

This lands directly on the gap this week's hacking disclosures expose: point-in-time checks alone can't catch behavior that builds up over many steps. Teams running agentic workflows should treat sequence-aware policy enforcement as a baseline control.