On Monday you ask an agent for a discovery brief. You want to know whether the AI meeting-summary idea your customers keep raising is a real product or just three people venting in the same six weeks.
The first pass is generic, so you push back. You tell it the brief should decide whether to put a team on this, not just describe the space. You make it commit to four interviews across four roles. You argue about whether the gap is capture, summarization, or distribution. You reject a section, ask for the questions to be sharper, trim the method down to three patterns or fewer. Twenty minutes in, you have something genuinely good: a brief that knows what it is trying to learn and how it will know when it has learned it.
Then you close the tab.
On Friday you want that brief again. You remember the four roles. You remember the line about three patterns or fewer. You do not remember which thread it was in. You scroll. You find a version, but it is the second draft, before you fixed the method. The good one is somewhere above it, between a tangent about Slack versus email and a completely unrelated question about a pricing page. You give up and ask the agent to regenerate it. The new brief is fine. It is not the one you spent twenty minutes making good.
The AI was not bad. The output had no home.
Two different things wearing the same clothes
There are two activities happening in that Monday session, and chat presents them as one.
The first is conversation. Negotiation, exploration, correction, iteration. You and the agent are figuring out what the brief should be. This is genuinely good work, and chat is genuinely good at it. The back-and-forth is the point. You want the false starts, the rejected section, the moment where you said "no, decide, do not describe." That texture is how you arrived at something worth keeping.
The second is the artifact. The brief itself. The thing that should survive the conversation that produced it.
Conversation is meant to be transient. You would not want to reread the negotiation; you would want the result of it. The artifact is meant to be durable. You will come back to it on Friday, and again next month when you decide whether the interviews changed your mind.
Chat treats both as the same kind of object: a message in a stream. The twenty minutes of negotiation and the one paragraph worth keeping get the same storage, the same retrieval, the same permanence, which is to say almost none. The conveyor belt does not distinguish between the part you were supposed to throw away and the part you were supposed to file.
The four ways it fails you
Once you see conversation and artifact as different things, the specific failures stop feeling like quirks and start looking structural.
There is no retrieval. A thread is searchable only by the words that happen to be in it, and the useful artifact is usually surrounded by words that are not in it. Your discovery brief is buried under a debate about delivery channels and a question about a landing page. Full-text search over a transcript is not the same as being able to ask for "the brief, the good version." You are searching a record of how you thought, not an index of what you made.
Nor is there structure. A thread is one axis: time. Everything is ordered by when it was said, which is exactly the wrong axis for work you want to reuse. The brief, the four transcripts, the synthesis of the three patterns, and the follow-up survey are one connected piece of research. In chat they are scattered across whichever sessions you happened to run them in, in whatever order the days fell. Nothing holds them together except your memory that they are related.
Durable and transient work are mixed together. This is the core of it. The keepers and the scratch live in the same place, in the same format, with no marker telling them apart. You cannot point at the thread and say "this part outlives the conversation." The signal is buried in the noise that produced it, and over time the noise wins, because there is always more of it.
And portability is weak. The brief exists, but it exists as a message inside one provider's interface. You cannot easily hand it to a teammate, point another agent at it, or open it next year. To move it you copy and paste, which strips whatever structure it had and quietly resets the clock on it being lost again.
None of these is a bug in a particular product. They follow from asking a stream to do the job of a filing cabinet. A stream is good at one thing: showing you what just happened. That is the opposite of what an artifact needs, which is to be findable long after it stopped being recent.
The graveyard fills up quietly
The unsettling part is that nothing announces the loss. There is no error, no deleted file, no warning that the good version is about to slip below the fold. The brief is still technically there. You could find it if you scrolled long enough and read carefully enough. You just never will, because the cost of recovering it is higher than the cost of regenerating it, and so you regenerate. Sometimes the loss is louder, as when a thread hits Claude's maximum conversation length and refuses to continue at all (a guide for that one). The quiet version costs more, because it never even registers as loss.
Multiply that by a month. The research brief. The launch page you refined three times. The recap format you finally got right. The customer-language summary you would kill to have in front of the next person who asks "what did we actually learn?" Each one was real for a moment, then became sediment. You have a vague memory that good work happened. You could not produce most of it on demand.
This is not a discipline problem. You are not failing to take notes. The work arrived already written, often better than you would have written it, and the place it arrived in had no way to keep it. The agent did its job. The container did not have a slot for the output.
The work needs somewhere to go
The honest conclusion is narrow, and I want to keep it narrow.
Chat is not the problem. Chat is the right place to think with an agent. Keep arguing with it there, keep rejecting drafts there, keep doing the twenty minutes of negotiation that turns a generic brief into a sharp one. The conversation belongs in the stream.
The problem is that the conversation is currently the only place the artifact lives too. The thing you made and the talking you did to make it share one grave, and the talking always buries the thing.
If agents are going to keep producing real artifacts, and they are, then those artifacts cannot keep being born and dying in the same scrollback. The brief, the transcripts, the synthesis, the survey: at some point the keepers have to leave the conversation and go somewhere built to hold them. Somewhere with retrieval instead of scrolling. Somewhere with structure instead of one long axis of time. Somewhere that knows the difference between the part you were supposed to throw away and the part you were supposed to keep.
What that somewhere looks like is a separate argument, one I make in A Workspace Is Not a Folder. This piece only insists on the smaller claim, the one Friday makes for me every week: the useful thing was real, and it deserved a better fate than the stream it was born in.