Thariq Shihipar wants people to use a Claude feature that, he says, almost nobody is using yet: the dashboard artifact. Say Claude is working on a long project, perhaps one organised as a kanban board, where tasks move across columns from "to do" to "done". Claude can keep that board's data in a database attached to the artifact. Several Claude agents can read and update it, and the board can send information back to them. Shihipar works on the Claude Code team at Anthropic, where he teaches people to use the coding agent and also does engineering work. He says he has "done a bad job" of explaining everything artifacts can do.
He made the case on Latent Space, speaking with hosts swyx and Vibhu in an episode published on September 29, 2026. The dashboard is his starting point for a larger claim. He thinks the single coding session, where one person types to one agent in one terminal on one computer, is coming apart. Its thinking, its work and its human interface are splitting into separate pieces.
From chat window to generated interface
In Anthropic's products, artifacts are pages Claude creates: live HTML or Markdown documents rather than chat replies. According to Anthropic's Claude Code documentation, an artifact keeps a permanent address. Updates appear at that same address, and supported comments can be sent back to Claude for more work. The company's help page says stored data survives later visits. It also separates personal storage, such as a journal, from shared storage, such as a leaderboard.
The discussion questioned how far this should go: which feedback belongs in the artifact and which in the chat with Claude, given that the "AGI-pilled" choice, the one that assumes the models will keep getting much more capable, could be to send everything through Claude. Shihipar disagreed: "I think the more AGI-pilled one is to go through the artifact." In the long run, he said, artifacts become the user's window into the harness, the software loop that runs the model, hands it tools and carries out its actions. People would comment on a live document of the plan and watch several agents work, with each artifact built for the job in progress. "On-the-fly interface for your harness is probably where things are headed," he said. He added that the infrastructure is not there yet.
Brain, hands and surface
The discussion then turned to why, if artifacts connect everything, all work shouldn't happen that way. Shihipar answered by separating three questions: where the intelligence runs, where the work happens and where the person looks at it.
Today, he said, Claude Code is mostly local. A person starts a session on their own machine, and it runs there. Users can turn on Remote Control or start Claude Code in the cloud. He described Remote Control as a way "to get some cloud behavior", but Anthropic's Remote Control documentation is more limited. It lets a browser or phone control a session that still runs on the user's machine, and the computer has to stay on. A cloud session is different because it has its own hosted environment.
Shihipar said Anthropic is moving toward a Claude that lives in the cloud. You message it, and it starts sessions wherever they are needed. That, he said, is roughly how Claude Tag already works. Over time the company plans to add "local hands": the ability for that cloud agent to use your own computer when it is online. The cloud agent could start many subagents, smaller agents given parts of the task, and those subagents could talk to each other. The artifact would display all of it.
That gives three layers. The surface is the artifact, hosted somewhere with its own database. The brain is the model's reasoning, running in the cloud, so "you don't have to worry about shutting off your computer." The hands carry out the work on a laptop, in a remote sandbox (an isolated environment for running code) or wherever the job needs. He called this "unpackaging" a Claude Code experience that now happens in one place.
Anthropic's engineers have used the same terms before. In an April 2026 engineering post titled "Decoupling the brain from the hands", they describe a hosted agent system built on the same split. A lasting log of the session sits apart from the process that runs the model, and both sit apart from the sandboxes that execute code. If the model process fails, it can restart from the log. A sandbox can be replaced on its own, and credentials can stay outside the environments where code runs.
What exists now and what is planned
Shihipar was clear that some of this is shipping and some is not. Claude Tag, an agent that works in a company's Slack channels, is what he calls Anthropic's "multiplayer product." Projects applies the same idea to Anthropic's own Claude apps. You message it and it runs Tag-style work, such as starting subagents. He said Projects "will start off single player" and then expand.
Anthropic's September 17 Projects announcement describes a main conversation that plans the work, hands out parallel threads and reviews what comes back. Each coding thread runs in its own cloud session with its own copy and branch of the code, so parallel changes still have to be merged and conflicts resolved. At launch, cloud execution was available to selected Pro and Max subscribers, and Anthropic called local execution a future addition. Shihipar said local hands are still being worked on.
He did not claim there will be one right setup. Some people use Remote Control heavily and some use Claude Code on the web. Inside Anthropic, he said, Claude Tag is used a lot. He recommends Tag for businesses and Projects for individuals who want "some of that niceness of tag" without the admin setup. People iterating on a product might stay in the Claude Code desktop app. Background jobs such as code review, security work or starting a pull request (a proposed code change) fit Tag better.
Work that happens in channels
Shihipar's own habits show what the multiplayer version looks like. He creates a Slack channel for each feature he works on. When he needs a privacy, security or legal review, he tags Anthropic's Legal team in that channel. He tells them that Claude, which has followed the whole feature, can answer their questions. Legal gets precise answers about what exactly is going into the code, "and I don't need to be in the loop," he said.
He described on-call incidents as "inherently multiplayer." Several people join, tag Claude and look for context, and Tag can connect to a company's existing alerts. He also gave a startup example. When a new prospect lands in the company database, Claude can research it and then tag the right salesperson with next steps. He credited Andrej Karpathy with describing this kind of tool as an "organizational harness."
Some of the most powerful uses may involve nobody asking at all. Toward the end of the discussion, the conversation turned to proactive work, where Claude acts on events instead of waiting to be tagged. Shihipar agreed: "have Claude proactively do it."
Whose access, whose memory
Shared agents raise hard questions about identity and access. In the discussion, Claude Tag's decision to act under its own identity rather than any one user's was called controversial. Shihipar said people have to organise Tag channels themselves, and that Tag "does a good job of sanding down" the complications. Imagine hands on several colleagues' computers, or your connectors alongside someone else's. (Connectors are logged-in links to services such as Google Docs, often made through MCP, a standard protocol for connecting AI agents to outside tools.) To reach Google Docs, he said, Tag can use a shared, company-configured connection, or a person's own credentials when the shared one lacks access.
Anthropic's September 24 announcement of personal connectors in channels spells out that split. One request can combine a channel's access to a code repository with the requester's own Google Drive. The personal permission covers only that request and does not become a shared credential for the channel. Any answer posted in the channel is visible to everyone there. Scheduled and unattended jobs still use connectors that administrators set up.
Memory raises a subtler issue. The discussion noted that a human coworker tagged into many projects naturally carries knowledge from one to the next. With Claude, it is unclear whether it should use what it learned elsewhere or stick to the channel in front of it. Anthropic's Claude Tag documentation says memory and resources are bounded by organisation, workspace and private channel, rather than automatically carrying every participant's access. Administrators can view and manage what is stored. Direct messages use the individual's account, while activity in shared channels belongs to the company's deployment.
Shihipar called the permissions problem a "tip of the iceberg" situation and said Anthropic has spent a great deal of time on it, partly because it protects Anthropic's own codebase. He listed edge cases. A Claude in one channel with its own permissions might leak data by messaging a different channel. Or it might use one person's connector and then send the result to someone else.
His sharpest example involved prompt injection, where text written by an outsider slips instructions to an AI system. Suppose a company has a public suggestions page that feeds into Slack through an automated hook. Someone submits a suggestion with hidden instructions. The agent reading it has broad access to company data, and now "you've exfiltrated your code base." External Slack channels shared with other companies are useful places for Claude too, he said, but they widen the exposure. The more central this organisational harness becomes to a company, the more there is below the surface.
Cost and adoption friction
Shihipar acknowledged that the new setup is not quite ready, partly because it is "more token-expensive." Tokens are the units of text that AI usage is billed by. When Claude manages subagents and reviews their output instead of a person doing that, it uses more of them, and so does writing to an artifact rather than replying in plain chat. "I don't actually think it's too much more," he said.
Setup is another hurdle. Claude Tag is harder to adopt than Claude Code because an administrator has to install it. Shihipar compared it to Claude Code's slow start: people took time to understand what it could do, and he said Tag is exciting once users reach "the magic moment." The conversation also brought up mixed results outside Anthropic. Some organisations like it, and many people say they don't see the difference or why they would use it. Shihipar conceded that Anthropic has "lots of tokens."
His answer was about price trends. When Claude Code came out, he said, nobody was used to paying more than $20 a month for AI, let alone $200 subscriptions. Opus 4 was a big, expensive model, while Opus 4.5 was "both great and cheap." He expects the intelligence of Anthropic's current Fable models to become cheaper and more plentiful too, and once it does, spending tokens on tools like Tag will make sense. A September 25 cost guide on Claude's developer blog explains why sticker prices can mislead. A task's real cost includes repeated model calls, failed attempts and a growing conversation that is resent every turn, so a model with a higher per-token price can sometimes finish a job for less.
For businesses, Shihipar's advice is not to wait. Making company data available to agents takes time, he said, so that work should start now, even if firms hold off on heavy spending until models get cheaper. For companies tempted to build their own version of Tag, he pointed back to the suggestion box: with an agent in the middle of a company's data, "the security is really, really important."