Claude Tag: Anthropic Puts an Agent in the Channel
Anthropic's new Claude Tag turns @Claude into a persistent, shared teammate inside Slack. Here's what it actually does, how it's governed, and why the form factor matters more than the feature list.
Anthropic launched Claude Tag today, and the one-line version is deceptively simple: type @Claude in a Slack channel and an agent joins the conversation as if it were another teammate. The interesting part isn't that an AI can now read your channel; plenty of bots already do. It's that this one is persistent, shared, and scoped, and that the unit of interaction is no longer a private chat window but a public thread the whole team can see.
I think that distinction is the whole story, so let me unpack it.
What it actually is
Claude Tag is an agentic instance of Claude that lives inside a Slack channel. You summon it the same way you'd loop in a colleague, by tagging it. From there it reads the surrounding conversation, breaks the request into stages, executes using whatever tools it's been granted, and posts its work back into the thread. The kinds of tasks Anthropic is demoing aren't toy examples: opening pull requests, running data analysis, and helping work through incidents.
It runs on Opus 4.8, and it's available today in beta for Claude Enterprise and Team customers. Anthropic has said it intends to bring the same form factor to platforms beyond Slack in the coming weeks. Notably, it replaces the older "Claude in Slack" app rather than sitting alongside it; that legacy integration is being retired on August 3, 2026, and administrators can opt into migration within a 30-day window.
Anthropic frames Claude Tag as the next step in the evolution of Claude Code rather than a new product line. That framing is backed by a number they're clearly proud of: the company says its internal version of Tag now writes roughly 65% of its product team's code.
The four behaviors that matter
Strip away the announcement language and there are four properties that actually change how the thing feels to use.
It's multiplayer. A single Claude instance lives in each channel, not one per person. If a colleague hands it a task and steps away, you can pick the conversation up where they left off. Slack's GM described the shift as making AI "multiplayer," and that's the right frame: the assistant becomes shared infrastructure rather than a private tab everyone keeps their own copy of.
It accumulates context. Instead of treating every mention as a cold start, Tag builds persistent knowledge of a channel's history and the organizational data it's allowed to see. The practical payoff is that you stop re-explaining the same background every time. The cost, which is worth naming, is that the assistant's usefulness becomes a direct function of how much of your organization it's been allowed to ingest.
It can be proactive. With "ambient" mode switched on, Tag doesn't wait to be summoned. It can flag relevant updates, follow up on unresolved threads, and surface things it thinks you should see. Anthropic's own example: Cat Wu, head of product for Claude Code, connected her Tag to Gmail so it watches her inbox and pings her in Slack when someone important writes, a place she says she's more responsive than her email.
It's scoped per channel. Each instance is its own identity. A Claude configured for a sales channel doesn't share memory or tool access with one in an engineering channel. This is the feature that makes the other three deployable in a real company rather than just a demo.
Governance is the actual product
The capabilities above are only as good as the controls around them, and Anthropic seems to know that the enterprise sale lives or dies here. Administrators define which channels, tools, and data sources each instance can touch. They can cap monthly token spend at both the organization and per-channel level. And every task Tag executes is auditable, including who asked for it.
That last point deserves emphasis. An always-on agent that can read messages, files, and code, and act through connected tools, is a meaningful expansion of the attack and accident surface inside a workspace. The presence of scoped identities, spend ceilings, and a full task log is what separates "we shipped an agent" from "we shipped an agent our security team will actually approve."
Why the form factor is the point
A lot of what Claude Tag does was technically possible before: you could already route coding tasks to Claude from Slack, and agents have been able to call tools for a while. What changed is the interface. Tagging an agent the way you'd tag a coworker collapses the distance between "discussing the work" and "doing the work." The bug report, the context, and the fix now live in the same thread, visible to the same people.
For anyone building agentic systems, that's the lesson worth stealing.
The hard part of these products is decreasingly the model and increasingly the surface you wrap it in: where the agent lives, what it remembers, who can see it work, and how tightly its reach can be bounded.
Claude Tag is a bet that the winning surface is the place teams already are, and that "shared and governed" beats "private and unconstrained" once you're inside a real organization.
Whether ambient proactivity turns out to be genuinely useful or quietly annoying is the open question I'd watch. An agent that surfaces the right thing at the right moment is a gift; one that interrupts a channel with low-signal nudges will get its ambient mode switched off within a week. That outcome won't be decided by the model. It'll be decided by how well the proactivity is tuned, which, fittingly, is exactly the kind of loop you can only get right in production.