AWS offers local, open source leash for agent harnesses

AWS has introduced an open source software library, called Dogwood Local Engine (DLE), designed to add policy control to AI agents. Much like a dog harness requires a leash to prevent the animal from running wild, DLE aims to keep AI agent tool calls within user-defined boundaries. The DLE is a Rust library that can be embedded into an agent's harness or gateway, issuing allow/deny verdicts for each tool call.

It does not enforce the decisions, but it checks the actions against policies defined using Dogwood, an open source governance language AWS introduced in August. DLE's key feature is its awareness of temporal conditions, tracking tool call events over time and persisting them to a disk, ensuring state retention even in case of system crashes.

AWS provides an example of using DLE to control coding agent Git pushes, defining a policy that only allows a push when the most recent test run has passed within the past 15 minutes. The DLE integrates seamlessly with Amazon Bedrock AgentCore, making policy enforcement directly embeddable into agent harnesses. AWS claims the added delay introduced by DLE is minimal, with evaluation time around 20 microseconds for 15-minute windows and increasing to about six milliseconds for 24-hour windows.

The library also prevents concurrent submission collisions by using a lock that allows only one event submission at a time, maintaining the lock until policy evaluation is complete. While there may be concerns about potential incorrect enforcement in cases where concurrent submissions occur, AWS aims to provide a solution to prevent such issues.

As AI agents become more autonomous and work for longer intervals, DLE offers a safeguard to regulate tool usage and prevent irreparable consequences.