The most useful person in the Slack thread may soon be the one nobody has to brief. Anthropic's Claude Tag lets anyone on a Claude Team or Enterprise plan type @Claude into a Slack channel or thread. Claude reads the conversation around it, pulls from the tools and data an administrator has connected, does the work and posts the result back in the thread. This week the head of Claude Code made the case that it has become his main way of shipping software.
"Tag writes >50% of my PRs every day," Boris Cherny, who leads Claude Code at Anthropic, wrote on X on September 25. "It also does ~100% of my data analysis, and fixes most of our product feedback + bugs." He argued that the product is a different kind of thing from the chat integrations that came before it: "It's not like a regular Slack bot. It is proactive, programmable, it has memory, it has access to your connectors."
The prompts Cherny shared are closer to instructions for a teammate than questions for a chatbot. In one, he tells Claude to mark threads in a channel with a checkmark emoji once they are resolved. In another, he asks it to try to reproduce every bug reported in a channel end to end by running the full app, then open a pull request with a fix and tag the right team for code review. For data work, he asks it to come up with roughly 100 hypotheses for an odd dataset, test each one in a workflow, spend "10m tokens or so digging very deep," and chart what it finds. He credited the newest models, Opus 5.5 and Fable 5.1, with giving the agent "strong judgement." Opus 5.5 was released on September 22.
How It Works
Claude Tag first shipped in beta on June 23 and replaced Anthropic's older Claude in Slack app. Anthropic's launch post says that when someone tags @Claude, it "break[s] its task down into stages" and works through them with whatever tools it has, then replies in the thread with what it made. Each channel shares a single Claude identity, so anyone can see what the agent is doing and take over a task where a colleague left it. Claude also builds up memory from the channels it belongs to. If an administrator allows it, Claude can learn from other channels and data sources as well, though it does not report from private channels. An optional "ambient" mode lets it flag information on its own and chase threads that have stalled.
Administrators decide which tools, data and codebases each Claude identity can reach, and in which channels. Anthropic says memories stay within those boundaries, so an agent set up for sales work will not carry what it knows into engineering channels. Admins can cap token spending for the whole organization and for individual channels, and they can see a log of everything Claude did and who asked for it. Users can also message Claude directly for private requests that rely on their own personal connectors.
Anthropic has been open about how much of its own work runs through the tool. At launch, the company said 65% of its product team's code is created by its internal version of Claude Tag. It said the same approach now handles product metrics, support tickets and root-cause analysis of bugs. "We see Claude Tag as an evolution of Claude Code," Cat Wu, Anthropic's head of product for Claude Code and Cowork, told Fortune in June. "Claude Code, Cowork, and chat are very single-player, whereas Claude Tag is built to be interactive and multiplayer."
On the same day as Cherny's post, Anthropic's developer account announced a smaller change for Claude Code users. When a five-hour usage limit arrives in the middle of a task, Claude Code "will now try to find a graceful stopping point" instead of stopping mid-edit. To finish up, it gets a small, fixed allowance taken from the user's weekly limit.
Why It Matters
Cherny's numbers come from the person who runs the product, so treat them as a best case, not a benchmark. They still point to where enterprise AI is heading. The competition is less about which chatbot a company licenses and more about where the agent sits and how much of the company it understands. Slack threads hold the bug reports, the half-finished decisions and the unwritten knowledge that agents normally lack. An agent that reads them and can act through connected repositories and databases starts to function like a coworker.
The timing makes the competition hard to miss. Microsoft on September 25 unveiled a rebuilt "New Copilot," which includes an always-on agent called Autopilot that has its own identity inside a company's Microsoft 365 tenant. A day earlier, the startup Ando came out of stealth with $20 million from Accel, Index Ventures and Emergence Capital. Ando is building a messaging platform where AI agents join channels, DMs and calls as members of the workspace. TechCrunch framed the pitch as a direct challenge to Slack. The three efforts share one premise: the chat window is becoming the control plane for agent work. They differ on who owns that window, which could be Slack, Microsoft or a new app built for agents from the start.
Anthropic's approach avoids that fight by running on Slack, where many of its customers already work. The trade-off is dependence on Salesforce's platform, which is pushing its own Slackbot as an AI assistant.
What to Watch
Three things to watch. First, whether Claude Tag leaves beta and expands past Slack; Anthropic has said it wants teams to be able to tag @Claude "in the many other places they work." Second, whether customers outside Anthropic report results anywhere near Cherny's, especially on code and data analysis, where autonomous pull requests and multi-million-token investigations will test both code-review practices and token budgets. Third, how IT leaders compare Anthropic's channel-scoped permissions with Microsoft's tenant-wide agent identities once they have to choose which agent gets access to the company's data.
“It's not like a regular Slack bot. It is proactive, programmable, it has memory, it has access to your connectors.”— Boris Cherny, Head of Claude Code, Anthropic