Not subscribed? Sign up to get it in your inbox every week.

Earlier this week, Anthropic shipped a teammate.

It's called Claude Tag, and it replaces the old Claude-in-Slack app with something closer to a full teammate. You add it to a channel, and anyone on the team can type @Claude to hand it work. It answers in the thread, keeps going while you move on, and remembers what it picks up from the channel so you stop re-explaining the same context every morning. Turn on its ambient mode and it speaks up on its own, flagging a stalled thread or chasing down a task everyone forgot. It runs on Opus 4.8, and Anthropic says 65% of its own product team's code now comes from an internal version of it.

I've spent enough of the last two years inside these tools to recognize the shape of the move, and Tag is the third act.

The first act was chat. You opened a window, asked a question, and pasted the answer back into your real work. The model never touched your environment. It was a smarter search box, and the whole loop ran through your copy-paste.

The second act was cowork, and it was single-player. Think Claude Cowork or Codex: the agent runs inside your repo and edits your files while you watch. This is where most of the leverage of the past year actually came from, because the agent stopped advising and started doing real work on your machine in a long-running session. But it was a private session: one human, one agent, behind a closed door. Your teammate never saw what your agent did, and the moment you shut the laptop the context died with it.

Tag is the third act. The headline is that it lives in Slack. The deeper change is that it's multiplayer: one shared agent the whole channel uses, instead of a private one that belongs to you alone.

There is one Claude in the channel, not one per person. Anyone can tag it, anyone can see what it's working on, and anyone can pick up the thread where the last person dropped it. Its memory belongs to the channel instead of to your session, so context compounds across the whole team rather than evaporating when you log off. The human-in-the-loop used to be you, alone, steering a single agent. Now it's the whole team steering one shared teammate, in the surface the team already lives in.

That changes what "in the loop" even means. In the cowork era, being in the loop meant sitting in the session with your hands on the wheel. With Tag the agent schedules its own work, breaks a project into stages, and grinds on it for hours or days before coming back to the channel with progress. The loop got longer and quieter. You're less the driver and more the team lead who handed off the work on Monday and reviews it on Thursday.

If you read this newsletter, you've watched me circle one idea for months: the company that wins with agents builds a central place to store its skills, its tools, and its context, then lets every agent draw on it. We covered the cleanest real version back in April. Ramp built an internal suite, Glass, where one engineer's Monday skill is running on the whole support team's queue by Tuesday, with more than three hundred shared skills versioned and reviewed like code. That was a system Ramp had to staff a team to build.

Tag is Anthropic selling a version of that system to anyone on a Team or Enterprise plan. The shared agent, the channel-scoped memory, the tool permissions, the per-channel spend limits, the activity log: that is the skeleton of the central agent system, productized. The build-versus-buy line I wrote about three weeks ago just moved again, because the central layer a company like Ramp paid engineers to build is becoming something you switch on.

That shift also raises the stakes on something I wrote about in May: every agent needs a manager. A private cowork session that drifts is your problem for an afternoon. A shared agent that drifts is wrong in front of the whole team, learning from every message in the channel, and quietly trusted by people who never checked its work. Multiplayer makes the upside collective and the failure collective with it. The ambient pings are useful right up until they are noise, and a teammate nobody owns becomes the group-chat member everyone mutes.

So the discipline has to scale with the surface: one name on the shared agent, a clear line on which channels it can read and what it can touch, and a standing review of what it has been doing while you weren't watching. Anthropic shipped the governance knobs for exactly this (separate identities per use case, per-channel permissions, token caps), which tells you they expect it to be managed like a hire.

Here's the test for this week. Pick the one channel where your team already does real work in the open, the place where the context actually lives. Put a single shared agent in it, give it a name and an owner, and watch what it learns over five days. The teams that have been encoding their skills all year will get a useful teammate on day one. The teams that haven't will meet a junior one, in front of everybody.

Chat asked, cowork did, and now the whole team is in the loop.

Till next time,
Chris