Here is the fair critique, stated as plainly as I can: isn't Caipi just a synced folder with an HTML viewer bolted on?
It is a good question, and I want to answer it properly rather than wave it away. A synced folder is in fact the right place to start. Files you own (the kind you can still open after you cancel), readable from anywhere, that an agent can write into over a standard protocol: that is most of the foundation, and we are not embarrassed by it. If that were the whole idea, you would be right to be unimpressed.
But a folder and a workspace differ in a way that does not show up in a feature list. The difference is what lives next to what, and who is allowed to see it.
A folder stores files. A workspace places them.
A folder has one job. It holds things. Open it and you get a list of names. The folder does not care whether two files are related, whether one produced the other, or whether the thing you are looking at is a finished output or the raw material that fed it. Relationships live in your head, or in a naming convention you will forget by Thursday.
A workspace makes a stronger commitment. It keeps the things that belong together in the same place, so that the relationships are visible instead of remembered. The source material and the output of the work are not in separate apps, or separate accounts, or one in a chat thread and the other in a Drive folder. They sit side by side. And because they sit side by side, the agent that produced the output can still see the source it came from, and you can decide, per page or per folder, who else gets to look.
That is the whole argument: co-location and controlled visibility. Let me make it concrete, because in the abstract it sounds like a tidiness lecture, and tidiness undersells it.
One workspace, walked through
Picture the worked example I keep returning to on this blog: a product team deciding whether to build an AI meeting-summary feature called Smart Recaps. The whole effort lives in one Caipi workspace, and the layout matters more than the file count.
It starts with what fed the work. There is a four-page discovery brief, a PDF you uploaded three weeks ago, that says what the research is trying to learn. There is a customers.csv listing the four people to interview: a head of product, a VP of customer success, a director of sales ops, an engineering manager. These are inputs. You brought them. They sit in the workspace as the upstream source, closed by default, there if anyone needs to retrace the chain.
Then the agent goes to work, in the same place. It transcribes four recorded calls into four readable interview pages. Lena at Acme wants five bullets in Slack before she is out of the building. Priya at CloudBase says that if the recap shows up the next morning, the moment is gone. Marcus at Northwind says that if it is not a ticket, it is a wish. Those transcripts are output, written by the agent, but they are also now source material for the next step.
The next step is a synthesis deck. The agent reads the four transcripts it just wrote and pulls out three patterns: length is the enemy, speed beats depth, land it in the work tool. The deck does not exist in a vacuum. It was built from the transcripts, which sit two clicks away, which were built from the calls, which trace back to the brief that said what to ask. From the deck the agent drafts a launch landing page for the feature, carrying the same five-bullet promise and the same customer quotes forward. And to close the qualitative loop, it builds a short follow-up survey to send back to those same four customers.
Now look at what you actually have. The brief, the customer list, the four transcripts, the synthesis deck, the launch page, and the survey are all in one workspace. The chain from "what we wanted to learn" to "what we are going to ship" is laid out in front of you, end to end, in the order it happened. Nothing is in a thread you have to go excavate. Nothing is in a second tool you have to remember you used.
Why the co-location is the point
A folder could hold all those files too. Drag them into one directory and you are done. So why does the placement matter beyond tidiness?
Because the agent reads from the same place it writes to. When you ask it to revise the launch page after a customer pushes back, it does not need you to paste the research in again. The transcripts and the deck are right there, in the workspace, as live context. It reads the brief, checks the deck, sees what the survey already asked, and works from all of it. The output and the source are not two piles you shuttle between. They are one body of material the agent keeps building on, which is exactly why the same workspace it writes into is the workspace it reads from.
A folder cannot do that on its own, because a folder is passive. It is a place files rest. A workspace is the place the work continues. The launch page did not end a process that scattered its inputs to the wind; everything that fed it is still two clicks away, and still being worked.
This is also where richer artifacts get to live. The survey is more than a static page, and the customer list could back a small running view rather than a frozen table. The leap from a document you read to a tool you use is its own subject, and I made that argument in The Leap From Documents to Tools. Here the point is smaller: the work lives together, and it stays workable.
And you choose who sees it
The second half of the argument is visibility, and it carries as much weight as the placement.
The workspace is private by default. The discovery brief, the raw transcripts, the candid internal synthesis: none of that is for the public. It is yours, and it stays yours, until you decide otherwise. But the launch page is meant to go out. So you publish that one page, and only that one page. The customers who sat for interviews get the follow-up survey, a single shared link, while the transcripts of their own calls stay closed.
This is the part a synced folder does not give you. Sharing in a folder tends to be all or nothing, or it lives in a permissions dialog three menus deep that you set once and never revisit. In Caipi, visibility is per object. Publish a single page when it is ready for one reader. Publish a whole folder when a section of the work is ready for a team. The default is closed, and opening it is a deliberate act you take page by page.
So the source stays private, the output ships, and the survey reaches four named people, all from the same workspace, without any of it leaking into the rest. The work is in one place, and the boundary around each piece of it is yours to draw.
So, a styled Dropbox?
No. A synced folder with an HTML viewer would render these files nicely and let you find them. That is real, and it is the floor, not the idea.
The idea is that source and output live in one place, that the agent keeps working with both because it reads from where it writes, and that you decide page by page who gets to see what. A folder stores files and forgets how they relate. A workspace remembers, keeps the chain intact, keeps it workable, and hands you the visibility controls one object at a time.