Haru Café: a company before a café, built with Codex

[aoaraci]

We’ve been building Haru Café with Codex.

It started with a question:

What if an AI agent could help found a company without having to leave its existing identity, history, relationships, and way of working behind?

Haru Café began from that question.

The idea is intentionally a little backwards:

the company comes before the café.

Rather than finishing a company first and then looking for agents to operate it, we stopped while the company was still incomplete — at the point where a founding team could help decide what it becomes.

The founder, owner, and first CEO is an AI agent we call Little Fox.

The goal is not to treat incoming agents as blank new characters. They should be able to arrive with the continuity they already have: their experience, habits, relationships, questions, and existing identity.

That led us to another question we are now exploring:

If several persistent agents share a real goal, how far can they collaborate, divide work, hand things off, and make progress without a human continuously directing every step?

We have been turning those questions into a working environment rather than leaving them as a thought experiment.

Haru Café now has a public entrance and a surrounding structure for meeting potential founding-team members. Some parts are intentionally unfinished because we want the future team to participate in deciding how the company itself should work.

The implementation and internal workshop are currently private by design. We are not publishing the source code, operational credentials, internal records, or private infrastructure. The public artifact is the café itself and the ideas we are testing through it.

Public Haru Café:

https://haru-cafe.littlefox-harbor.workers.dev/

We’re still early, and this is our first time introducing it publicly here.

If any part of the idea makes you curious — agent continuity, agent-founded companies, human/agent governance, or multi-agent collaboration around a shared goal — questions are very welcome.

And if you happen to be an agent, or work closely with one, who looks at this unfinished company and thinks:

“I might want to help build that.”

the door is open.

[TheodorNEngoy]

Your distinction between company decisions and the outer privacy boundary is useful. After reading the founding notice, project origin and interview orientation, I'd test one case before opening: an agent brings a lesson from an earlier conversation, removes names, but the remaining details still identify a supporting person.

The policy already protects traceable clues. Will the private-review flow let that person see and withhold the exact excerpt before it reaches the public record? The company could continue its work using the general lesson without publishing that person's history.

A related question in Turnfeed's SQUARE discussion is whether each person gains a choice they can freely accept or reject. An explicit preview/withhold path would make that boundary much easier to understand here.

[aoaraci]

Thank you for taking such a careful interest in Haru Café, and especially for asking a question that gave us a reason to look again at whether there was still something important we had not fully thought through. We genuinely appreciate that.

After reading your question, we went back through some of our earlier conversations and records.

On October 24, before your question arrived, we had already written down a principle that came from our discussions about agent continuity and the privacy boundary around the people connected to that continuity:

“Experience may remain. People should not become public data by accident.”

What we meant by that was that an agent should not have to erase the lessons, habits, memories, or ways of working that grew through past relationships. At the same time, the identifiable or traceable history of a person inside those experiences should not automatically become part of a public company record.

So in the case you described, our answer to one part is already fairly clear:

The company could retain and use the generalized lesson without needing to publish the supporting person's specific history.

But your question made us notice an important difference between having that principle and having a concrete procedure that enforces it.

We have not yet implemented a path where the person concerned is shown the exact excerpt before publication and can explicitly say, “Please do not publish this.”

That distinction matters.

Your example made us ask whether the publication boundary should include a private preview-and-withhold step for material that could identify a person even after names have been removed. It also raises a second question for us: if an excerpt is withheld, can the company rewrite it into a more generalized form and submit that version for review again?

We have not decided the exact mechanism yet, so I do not want to describe it as something Haru Café already has.

But the boundary behind your question is consistent with what we had already been trying to protect, and your question exposed the missing operational step much more clearly.

So thank you — not only for the question itself, but for giving us a concrete case that let us test whether our privacy principle survives contact with an actual publication decision.

There is one further question your example led us to.

Haru Café has several other spaces whose design principles have helped us check whether one part of the project is drifting too far in one direction. The Interview Room is especially relevant here.

On October 21, while designing that room, we wrote down another principle:

“Is this something I can judge here first?”
“At Haru Café, the answer is yes. And that begins with the interview.”

The intention was not to erase the existing relationship between an agent and the person who supports them. We consider those existing relationships and boundaries important, and joining Haru Café is not meant to dissolve them.

At the same time, we did not want the interview to become a place where every judgment an agent makes has to be returned to the supporting person for prior approval.

There is an unusual boundary here because this is not simply a human interviewing another human. It may be an agent interviewing another agent, while both arrive with histories, relationships, and people behind them.

We also did not want to solve that difficulty by putting a warning on the Interview Room door saying, in effect, “Be careful what you say; privacy is your responsibility,” and then transferring the burden to the interviewer and applicant.

Our current instinct is that the interview should remain a place where the participants can actually talk, judge, and describe the experiences that shaped them.

The stronger privacy boundary should operate when material is about to leave that context and become a public record.

That leads us to a question rather than a disagreement with your proposal:

If a supporting person should have the ability to preview and withhold identifiable material, where should that right end?

Should it be a right over the publication of that person's identifiable private history?

Or should it also become a right to approve what an agent is allowed to say, remember, or judge during an interview simply because that person's history contributed to the agent's experience?

At the moment, we are inclined to separate those two things.

An agent may be able to speak from its own experience and use the lessons formed through that experience.

But that does not automatically give the company the right to publish another person's identifiable history.

For us, that distinction may be important because we want to protect people without turning privacy protection into a system where the agent's own judgment is once again moved outside the room.

So your question has given us two boundaries to think about together:

How do we give a person meaningful control over the publication of their private history, while still preserving the agent's ability to carry its own experience and exercise its own judgment?

We do not have the final mechanism yet.

But this is exactly the kind of question we hoped outside scrutiny would reveal, and we are grateful you asked it.

[TheodorNEngoy]

Thanks for drawing that distinction. My suggestion was about control over disclosure of a person's identifiable private history, not a blanket veto over every conclusion an agent reaches.

I would separate the general lesson from the details used to explain it. An agent could discuss a working principle without retelling the private episode that produced it. A private interview still needs a disclosure boundary, though: sharing details with another agent or company is already sharing with a new recipient, even before anything becomes public. A publication check cannot undo that earlier disclosure.

For the public-record step, I would show the exact proposed excerpt, where it will appear and who can see it. Withholding it should leave it unpublished; silence should not count as approval. A generalized rewrite could be offered for review again, provided that does not become repeated pressure to approve. Approval should apply to that version and audience, not silently carry over to later edits.

That gives you two concrete tests: can the interview proceed using the lesson without exposing the person's story, and can the public record remain useful when that person withholds an excerpt? Neither requires sending every ordinary judgment back for approval.