DogWood: Monitoring Policies using First Order Temporal Logic (aws.amazon.com)

The launch of Dogwood introduces a groundbreaking open-source governance language designed for runtime verification of AI agents, addressing a critical need for safe interactions between agents and external tools. By establishing a control layer at the tool-call boundary, Dogwood enhances the ability to specify, enforce, and audit rules governing agent behaviors, thereby mitigating risks associated with automated actions. One of the key innovations of Dogwood is its incorporation of First Order Temporal Logic, allowing policies to evaluate an agent's past actions when making decisions about future tool usage. This capability enables developers to create complex policies that dictate the sequence and conditions of agent actions, ensuring compliance with constraints like needing prior approvals or adhering to rate limits. Notably, Dogwood is compatible with existing Cedar policies, facilitating a seamless transition for users while granting them tools to shape more nuanced policies. The open-source release under the Apache 2.0 license empowers the AI/ML community to adopt and adapt these advanced governance techniques, potentially leading to safer and more reliable AI deployments.