A clear job
Tell it what to do, what done means, and where the job ends.
AI agents for real work
Dutch Door lets agents do more on their own. Each job gets only the access it needs. Risky actions pause before they cost you.
Made for many AI agents—not just one model.
Any AI agent
Wants to take an action
Dutch Door
Checks the job, key, and risk
Within limits
Moves
Needs approval
Waits for you
Email · files · code · data · money · browsers
The way companies already work
You would not give every worker every key. Give the agent a job, the right key, and a door it cannot open alone.
Tell it what to do, what done means, and where the job ends.
Let it use only the tools and data this job needs.
Make it stop before it sends, pays, deletes, or goes live.
The rules live in the workplace—not just in a prompt.
See the full workplace mapWhy Dutch Door
I fixed the prompt. I wrote the rule. I added docs. The same kind of mistake came back. Each quick fix left more for the next agent to remember.
Dutch Door began with a new question: what if each mistake made the system safer next time?
The risk is not only AI. Decision fatigue, a forgotten policy, an honest bad call, or deliberate misuse can send a person through the same open door.
Read the founder storyWhat the research shows
People get tired. Agents can miss rules. Bad actors look for a way around them. A wide key can still open every door.
That is why high-risk actions need a limit outside both the person and the agent.
See the studies and real cases
Mistake-proof by design
A workshop Dutch door lets valuable work pass through the open top while the closed half protects the busy floor. Dutch Door does the same for people and AI: useful work moves; risky actions pause at the threshold.
Go one level deeperPrivate pilot
Show us where an agent saves time—and where one wrong action could hurt. We are starting with one workflow at a time.