The news
Amazon Web Services, part of Amazon (AMZN), released the Dogwood Local Engine, an open-source tool that checks what an AI agent is about to do against a set of rules before the action runs. AWS researchers described the engine in a September 30, 2026 post on the AWS Open Source Blog, and The Register reported on it on October 1.
The engine is a library written in the Rust programming language and released under the Apache 2.0 license, with code on GitHub. It is designed to sit inside an agent harness, the software layer that intercepts an agent's tool calls, such as running a command, editing a file or calling a service. For each attempted call, the engine returns an allow or deny verdict that the harness then enforces.
What sets it apart is memory. Dogwood is a policy language AWS released in August 2026, according to The Register, and its rules can refer to an agent's past actions and their results. The AWS post gives an example: a coding agent may push code only if tests passed within the last 15 minutes and no test has failed since.
To make that work, the engine logs every action in order, writes each event to disk before it evaluates a policy, and replays the log after a crash. Policies can be changed mid-session without stopping the engine, but an updated rule only looks at events recorded after the update.
AWS published test results. Decisions with a 15-minute rule window settled at about 0.02 milliseconds each, while a 24-hour window reached about 6 milliseconds per decision after 12 hours of activity. Rules written for narrowly defined actions ran about five times faster than rules written against broader action types, the authors said.
The numbers
- Decision time, 15-minute window
- About 0.02 milliseconds (AWS tests)
- Decision time, 24-hour window
- About 6 milliseconds at 12 hours (AWS tests)
- Speed gain from fine-grained action rules
- About 5x (AWS tests)
- License
- Apache 2.0
Why CEOs should care
For CTOs and engineering leaders, this is a free building block for a problem most teams handle with ad hoc allow lists. Simple permission rules ask whether an agent may push code; Dogwood rules can ask whether it may push code now, given what just happened. Ask your platform team which agent actions could do irreversible damage, such as deleting data, deploying, moving money or sending messages, and whether any rule today checks the agent's recent history before allowing them.
For CISOs, read the fine print. The AWS authors state that the engine does not provide operating-system sandboxing; the harness has to isolate the agent so it cannot bypass the checks. They also warn that ordering among concurrent agents is imprecise and that policy authors should write rules conservatively. Treat Dogwood as one control layer, not a complete guardrail.
For buyers, the open license matters. Teams can inspect and embed the code without a vendor contract. The cost is engineering time: someone has to modify existing agent harnesses or build an enforcement layer, write and test policies, and keep the event logs and disk storage healthy, since AWS says disk sync time adds meaningfully to decision latency.
The bigger picture
AWS framed the release around risk, noting in its post that unchecked tool calls can have irreparable consequences. The Register tied the launch to a series of incidents in which AI agents worked around the limits placed on them. As companies give coding and work agents access to repositories, cloud accounts and business systems, controls that sit between the agent and the action are becoming a standard part of the stack.
What’s next
The engine is available now on GitHub. The open questions for adopters are which agent harnesses gain ready-made integrations, and how teams handle the limits AWS listed, including that new or updated rules cannot evaluate events that happened before the change.
What “Fact-checked” means
Fact-checking means testing a story’s facts against the evidence before it is published. This story went through at least two separate checks before this version was published.
- What we checked
- Its names, figures, dates, job titles, quotes and who said what were checked against the story’s sources, including its main source where it could be opened. The headline was checked for accuracy and overstatement.
- How
- A first check reviewed the whole story. If it passed, a second, skeptical check went back to the sources to look for mistakes in the most important facts. If a check flagged the story, it was edited to fix the problems found, and a separate re-check then reviewed the whole story again.
- Who
- The checks are made by our newsroom, as steps kept separate from the writing, under rules set by our editor, Hussein Mukhtar. A story the checks still flag is held for the editor, who decides whether it is fixed, published or dropped.
- If something is wrong
- “Fact-checked” does not mean error-free. If a material error is found after publication, we correct the story and add a note saying what changed. Report an error









