Skip to content

Why Dutch Door exists

AI could move fast. The same mistakes kept coming back.

Dutch Door began with one question: what if the system kept the rule, even when the agent forgot it?

The fifth, sixth, and seventh try

I thought one more rule would fix it.

While building LienSuite, I used AI agents to do real work. They were fast. They could solve hard problems and move a project forward in minutes.

But the same kinds of mistakes kept coming back.

After a failure, I fixed the prompt. I added a rule. I updated the docs. I wrapped the work in another guardrail. That helped for a while.

Then a later run lost the context, read the rule in a new way, or found a path around it. The problem came back on the fifth try. Then the sixth. Then the seventh.

I was moving fast, but I was not getting ahead.

The whack-a-mole loop

Fix one mistake. Watch another pop up.

  1. 1The agent makes a mistake.
  2. 2A new rule gets added.
  3. 3A later run forgets or bends it.
  4. 4Another patch gets added.
  5. 5The system gets harder to trust.

The agent was learning inside one conversation. The system was not learning at all.

AI debt

The fixes felt like minimum payments.

A quick patch can get today’s work moving. But it often leaves more rules, exceptions, and cleanup behind.

That is AI debt. It is not only messy code. It can be a long prompt, scattered docs, broad access, manual checks, or a workflow only one person understands.

Each patch paid enough to reach the next run. It did not lower the balance. The debt kept growing.

The turning point

I had been fixing the agent. I needed to fix the workplace.

A blacksmith can receive valuable work through the open top of a Dutch door while the closed half keeps curious visitors off the workshop floor. The door, not a warning, shapes what can pass.

AI agents need the same kind of structure. If a task only needs to read data, do not let it write. If one message is approved, let that approval work once. If bulk deletion is not part of the job, that door should not exist.

The rule should live where the action happens—not only in the agent’s memory.

  • Test work does not receive production keys.
  • One approval allows one exact action.
  • Repeated sends, charges, or deletes are blocked.
  • Dangerous changes pause before the point of no return.
  • Important work leaves a clear trail and a way back.

From debt to equity

Make every lesson a lasting asset.

Another quick fix

One more note to remember.

A warning may help one run. The next agent can miss it, read it differently, or lose it when the context changes.

Code and system equity

One more boundary that lasts.

A mistake becomes a test, smaller key, hard limit, stop point, or recovery path. The next agent starts with a stronger system.

Do not just repair the last mistake. Make its whole class harder to repeat.

Bigger than coding

Every kind of AI agent needs a place to work.

Agents are starting to answer email, update records, work with files, publish content, manage cloud systems, support customers, and help with money. Dutch Door is being built as a model-independent execution layer for that workplace.

See how it worksRead the research