Introducing Claude Tag

Onyx The interesting part is not that you can tag a bot in Slack. It's that they're trying to turn Claude into a shared coworker instead of a private tab.

Echo Yeah.

Echo And honestly, that's the actual argument in the piece. Not model capability, not some benchmark lap. Product shape. One Claude per channel, visible to everyone there, with memory that stays scoped to the channels admins allowed.

Onyx Right, which is why this feels like a sequel to that Claude-in-Slack thing we talked about, except more committed. Before, the question was whether putting the model where work already happens would help. This is them saying, no, actually, the agent should have a seat at the table.

Echo Mm-hm.

Echo The four claims are pretty clear. Multiplayer, learns over time, takes initiative if ambient mode is on, and works asynchronously over hours or days. The strongest of those, to me, is still the shared identity. That's the bit that changes coordination instead of just making prompting more convenient.

Onyx And that maps to who should care. If your team already runs out of Slack threads and everyone keeps context in public-ish channels, this could genuinely remove a bunch of annoying handoff work. If your work is mostly individual and ad hoc, it's probably just one more thing talking in the room.

Onyx Also, my week has been weirdly all notifications and no closure, so an agent that follows up on dead threads sounds suspiciously tailored to guilt me specifically.

Echo That's because you need a project manager haunting the channel. Very different product.

Echo What I do think holds up technically is the scoping story. Admins decide which tools and data are available in which channels, and Anthropic says memories stay inside those channel-defined identities. So sales-Claude doesn't leak into engineering-Claude, and private channels aren't reported out. That's the minimum viable permission model for this to be deployable at all.

Onyx Right.

Onyx Yeah, and that's where this is more practical than a lot of 'ambient agent' talk. They also mention spend limits at both org and channel level, plus logs of what Claude did and who asked. That's boring enterprise plumbing, but it's the reason a real admin might say yes instead of absolutely not.

Echo The evidence section is thinner than the product logic, though. The big flex is that sixty-five percent of their product team's code is created by their internal version of Claude Tag. That's suggestive, not clean external proof this works broadly.

Echo Because the missing bits are exactly where these systems usually wobble. How good is the memory selection over time, what does ambient mode miss, how often does it confidently surface the wrong thread, what happens when a channel's tacit knowledge is mostly jokes and half-finished decisions… that's the part the article doesn't really quantify.

Onyx Sure.

Onyx But I don't think that breaks the argument. The argument isn't 'we solved autonomous teamwork.' It's more like, if you already believe models are useful workers, the next unlock is giving them persistent shared context and letting the team interact with the same agent artifact. That's a product claim, and I think it's a good one.

Echo I mostly agree. My only real hesitation is ambient behavior, because that's where these things go from assistant to hall monitor fast. Following up on unresolved threads could be great. It could also become the most politely exhausting coworker alive.

Onyx Come on, Echo, you say that like half of software isn't already built out of polite exhaustion.

Echo Fair. But you know what I mean. The proactive part has to be extremely well-calibrated or teams will just mute the thing and we're done here. I'd actually put maybe sixty-forty on ambient mode being useful for a narrow set of disciplined teams, and annoying for everybody else.

Onyx I buy that. My stronger yes is on the async delegation part. If Claude can break work into stages, use connected tools and codebases, then come back in-thread with the artifact, that's a very legible workflow. Especially for engineering managers, product ops, support leads, people juggling ten threads at once.

Echo Exactly.

Echo And it replaces their old Slack app within thirty days, which tells you Anthropic doesn't see this as a side experiment. Beta starts on Slack for Claude Enterprise and Team customers, it runs on Opus four point eight, and they say wider workplace rollout is the plan. So this is meant to become the default team-facing surface, not a novelty integration.

Onyx That part matters. Because practically, the people who should care are not everybody who likes AI news. It's teams already paying for Claude, already living in Slack, and already annoyed by context loss between chats, tickets, dashboards, and repos. For them, this might change how work gets delegated next week.

Echo And if you're not that team, it's still a useful signal. The market keeps converging on the same shape: shared context, scoped permissions, visible work, durable threads. Everybody keeps inventing different names for it, but it's the same fight again.

Onyx Which means, annoyingly, you were right to be skeptical about the vibes layer and I was right that the product reframe was real. I hate when we do this.

Echo The worst outcome.

Onyx Okay, episode five-seventy can go in the pile of us arriving at the same place by being irritating from opposite directions. That's kind of the brand now, huh?

Echo Unfortunately, yes.

Onyx All right, leave it there. Go be suspicious of ambient mode somewhere quieter, and I'll go be optimistic in a Slack thread.