Andrej Karpathy posted a long thread on X this week about Claude Tag, Anthropic’s new Slack integration. Not a product announcement from Anthropic itself, but his own reading of it: this is, in his words, “the third major redesign of LLM UI/UX.” Big claim for what looks, at first glance, like a Slack bot. I read the thread twice, and I think he is partly right, and partly getting ahead of himself.
What Karpathy says
His argument, briefly. The first way we used an LLM was as a website: you go somewhere, log in, ask a question, close the tab. The second was the app: a program that lives on your own machine, closer to your files and your workflow. Claude Tag, he says, is the third form: “a self-contained, persistent, asynchronous entity with org-wide tools and context, working alongside teams of humans.” Not a place you visit, but an entity that is already present, in the channel where you were working anyway.
Concretely: you tag Claude into a Slack channel, it gets access to the tools and data you allow, and it stays there. It remembers what happened in that channel, it can pick up tasks while you focus on something else, and it reports back when it is done. Not a chat window you open and close, but a teammate that sticks around.
Karpathy is honest about the cost of that: “once you do all of the under the hood engineering work to make this ‘just work’ (e.g. across tools, integrations, compute environments, memory, security, etc.), Claude basically joins the team in a seamless way.” That parenthetical is, to me, the most interesting part of the entire thread, more on that below. He closes with a line I recognize: “it really takes a while to wrap your head around it, but it works and it is awesome.”
What I make of it
Two things stand out, and not both of them positive.
The “paradigm” mostly lives in that parenthetical, not in the flashy claim. Karpathy names tools, integrations, compute, memory and security in passing, between brackets. But that is exactly the work that determines whether this ever becomes more than a chatbot with a longer memory. I see the same pattern in my own work at SVB: the innovation is rarely in the idea “let an AI join the team,” that step is conceptually cheap. The innovation is in who gets access to which data, how you keep that auditable, and what happens when it goes wrong. A Slack bot that can reach everything is not a paradigm, it is an attack surface. A Slack bot that can reach exactly the right things, with logging and boundaries that actually hold, that is the real work. Karpathy acknowledges this, but his tweet gets remembered for the first half of that sentence, not the second.
“Persistent, org-wide entity” is exactly the kind of claim that makes me frown, coming from someone who works on digital trust and the EUDI Wallet. An AI that builds channel memory over weeks and months, that watches multiple teams at once, that schedules its own tasks: that is not just a UX improvement, it is a new kind of actor in your organization with its own memory and its own reach. Who decides how long that memory sticks around? Who sees what it “knows” about a sensitive project? Anthropic’s own materials describe this as scoped per channel, with permissions, and that is the right answer. But the question deserves more than a parenthetical, especially if this pattern spreads beyond Anthropic’s own product.
That said, the core of his observation holds. I now use Claude myself more as something I delegate work to than something I ask questions. That is a real difference in how the interaction feels, and it is not unreasonable to give it a name. I just would not call it a paradigm until the parenthetical is solved. I would call it a promise that stands or falls on unglamorous infrastructure work.
What I take from it
Not the hype, but the question underneath it: if AI becomes a teammate instead of a tool, the question shifts from “can it do this?” to “is it allowed to, and who decided that?” That is the conversation I would rather have now than after the fact.
This piece also appeared in Dutch: Karpathy noemt Claude Tag de derde grote redesign van LLM UI/UX.