AWS has introduced an open source library called Dogwood Local Engine (DLE) designed to restrict AI agent behavior by enforcing user-defined temporal policies on tool calls.

The library functions as a policy checkpoint within an agent's execution flow, evaluating each attempted action against rules established by developers. DLE does not execute enforcement itself but returns allow or deny decisions based on policies written in Dogwood, an open source governance language released by AWS in August.

Policies can incorporate time-based conditions, such as requiring a successful test run within the preceding 15 minutes before permitting a code push. DLE logs tool call events sequentially and maintains state persistence across system restarts, enabling temporal tracking of agent activity.

According to AWS, evaluation latency remains minimal, measuring approximately 20 microseconds for short workflows and rising to six milliseconds for extended 24-hour windows in internal tests. The system employs a locking mechanism to prevent concurrent policy evaluations, ensuring sequential processing of event submissions.

AWS emphasizes DLE's role in mitigating risks associated with autonomous AI agents that may execute unintended actions when operating without direct oversight. The company notes that as agents gain extended autonomy and tool access, such regulatory mechanisms become increasingly necessary to prevent potentially irreversible outcomes.

The library is available through AWS's GitHub repository, where developers can integrate DLE into agent frameworks or gateways to implement granular control over AI-driven operations.

AWS has not specified a timeline for broader integration of DLE into its commercial services, and the company has not provided further details on potential failure scenarios involving concurrent policy violations.