AWS Dogwood: AI Agent Tool Call Governance Solved

AWS has launched Dogwood, an open-source policy language designed to govern sequences of AI agent tool calls. This new tool aims to fix situations where an AI agent's actions might be valid individually, but incorrect when viewed as part of a larger process. Dogwood builds on Cedar, an existing open-source authorization language. While Cedar evaluates each action in isolation, Dogwood can consider earlier events and the order of actions. This means it can factor in things like whether a refund was previously approved or if an agent has already processed a certain amount in a given timeframe. AWS has also added Dogwood support to its Amazon Bedrock AgentCore Policy service. This allows developers to control which tools an agent can call and under what conditions, especially within longer workflows. The language and its reference implementation are now available under the Apache 2.0 license. The bottom line: Dogwood helps ensure AI agents make decisions that are not just individually correct, but also appropriate within the full context of their actions.