An annoying assistant is worse than a useless one. That’s the whole point of the new update for Claude Tag, Anthropic’s @Claude in Slack. It’s meant to get better at keeping quiet – and at speaking up at the right moment.
Before: one classifier, one message
Until now a small classifier decided whether Claude should reply on its own. It looked at each new message in isolation and made a yes-or-no call. Anthropic shows the problem with a neat example: two engineers chasing the same bug from opposite ends. Priya has a theory, Devon has the matching log data. Neither message is addressed to Claude, neither asks for help. Read one at a time, the classifier correctly does nothing – twice. Read together, there’s obviously work sitting right there.
Now: the whole channel, plus memory
The classifier is gone. Claude reads along instead. It now uses context from across the channel, its own memory, and the standing instructions you’ve given it. From that it picks one of four moves: reply inline when the answer is short, open a thread for the deeper work, route the message into a workstream it already has open – or say nothing.
In the bug example, Claude picks the second move without any @-mention: it sees Priya’s hypothesis and Devon’s evidence, opens a thread with the investigation already running, and pulls both in. According to Anthropic, that makes Claude roughly 30% better at judging when to step in and when not to. And the first response comes faster – it acknowledges you in seconds instead of starting up silently in the background.
You steer it in plain language: “Only respond here when someone tags you” or “Feel free to jump in on anything about the deploy pipeline.” In a channel where Claude keeps concluding, message after message, that it has nothing to add, it goes to sleep. An @-mention wakes it instantly.
My take
The interesting part isn’t the 30%, it’s the attitude behind it. Anthropic deliberately builds this thing to stay quiet most of the time. That’s the hard direction. Anyone can make an agent talk constantly. Building one that knows when it would be in the way – that’s the real work.
That the extra context doesn’t count toward your usage limits rounds it off. Otherwise you’d be paying, channel by channel, for Claude to pay attention. It’s live today for Teams and Enterprise customers.
Sources: