Amazon Web Services (AWS) has released an open-source policy language that can control not only individual actions of AI agents but also the sequences in which multiple actions occur.
AWS said on the 13th that it released a policy language, "Dogwood," that can validate continuous task flows based on an AI agent's past action history.
AWS applied a "time-based policy" built on Dogwood to the policy features of Amazon Bedrock Agents Core. While existing policies judged individual tool calls independently, time-based policies consider the agent's prior actions along with the current request to decide whether to allow the next action.
For example, it can block behaviors such as repeatedly making payments under the approval limit that cumulatively exceed the overall budget, or moving to the next step without going through a specific task. It can also require human approval before important tasks. Policies are enforced at the gateway layer, not in the agent code. AWS said this prevents agents from bypassing policies due to prompt injection or code errors.
Dogwood is compatible with the existing open-source policy language "Cedar" and is based on temporal logic that handles time and order. Released under the Apache 2.0 license, it can be used without rewriting existing policies.
AWS also added a "rate limiting" feature to the Agents Core gateway to cap AI usage per user. Usage can be managed by user, team, tool, and model based on metrics such as requests per minute, tokens processed per minute, and concurrent consolidation.
AWS said the two features can be applied without redesigning existing AI agents.