Andrej Karpathy, quote-tweeting the Claude Tag launch, called it the third major redesign of LLM UI/UX. The first paradigm was the LLM as a website you went to. The second was an app you downloaded. The third, in his words, is "a self-contained, persistent, asynchronous entity with org-wide tools and context, working alongside teams of humans."
Within a few days of each other, Anthropic put @Claude in Slack and Atlassian put @Jira there too. Then you have agent frameworks like Eve and Flue that let you easily connect your agent to multiple channels, Slack included.
We're shifting to agents as teammates. We'll see what that means for lock-in and costs, but there feels like a lot of potential for bringing agent power to teams. It could be a great opportunity to get more people on teams working with agentic power instead of agents locked behind the power users machines.
It could also be a great way to burn tokens and compute on someone else's infrastructure while racking up a hefty bill.
Eventually maybe we’ll even get Slack channels with agents talking to agents. For a great look at how that might work out, check out Season 2 of Shell Game.
Let’s get on to the links.
💬 Platform
Introducing Claude Tag
Anthropic put @Claude in Slack as a multiplayer teammate. Add it to a channel, connect your tools and codebases, and anyone can tag it in to delegate work that runs async on Anthropic's infrastructure. It remembers the channels it sits in and can act on its own, and Anthropic says 65% of its product team's code already comes from their internal version.
🎫 Workflow
@Jira: create work items from any Slack conversation + Deeplink to your AI coding tool from Jira
Atlassian made Jira agentic from two directions. Tag @Jira in any Slack thread to turn the conversation into a structured work item and hand it to a teammate or an agent like Cursor or Claude Code. And from any Jira work item you can now deeplink straight into your coding tool with the ticket's full context pre-loaded through Atlassian MCP, so the agent starts informed instead of cold.
Anthropic studied ~400,000 Claude Code sessions and found a clear division of labor: people make about 70% of the planning decisions while Claude makes about 80% of the execution. The real signal for non-engineers is that domain expertise, not coding skill, is what predicts success.
Shopify's internal Quick platform lets anyone at Shopify drop a folder of HTML and get a secure, employee-only URL, with database, AI, and websockets each one API call away. Over half the company has built at least one of its 50,000+ sites, all running on a single $200/month VM. Its one way to solve the issue of how to securely share internal tooling made by AI.
🧩 Components
New chat-interface components
shadcn shipped a set of composable, mostly headless components for building chat interfaces, starting with the conversation layer: streaming, scrolling, bubbles, attachments, and a MessageScroller that owns the fiddly scroll behavior while you bring your own data and model. The opinionated core is a rule worth stealing: never move the reader against their intent, so auto-scroll is never the default.
Thanks for reading,
Jason
