Claude Tag reads the whole channel: scope your Slack agent

The update Anthropic shipped to Claude Tag looks small and isn't. Its Slack agent now reads across an entire conversation instead of judging each message on its own, which Anthropic says makes it roughly 30% better at deciding when to jump in unprompted — and, more usefully, when to say nothing. The line that should actually stop an operator is the billing footnote underneath it.

What actually happened

Claude Tag is Anthropic's Slack agent, launched in beta for Claude Enterprise and Team as a replacement for the older Claude in Slack app. One Claude per channel, working in public view, picking up where the last person left off.

The change, reported by VentureBeat: Claude now holds the surrounding conversation as context and picks one of four moves — reply inline, open a thread and do deeper work, route the message into an existing workstream, or stay silent. Anthropic's own framing for the restraint is blunt: an annoying agent is worse than an unhelpful one, so Claude goes dormant in channels where it has nothing to add.

Then the footnote. That expanded channel context does not count toward usage or spend limits on any plan — for now. Anthropic's Scott White told VentureBeat the company is in "active experimentation and co-design with customers" and declined to commit to how it gets metered later. On the admin side you get budget caps tied to agent identities, role-based access groups, and model entitlements so different teams run different cost-performance profiles.

Why an always-on Slack agent matters for your business

Unbilled context is a preview of a bill. Something is reading every message in every channel Claude sits in, and today Anthropic absorbs that cost. Inventory which channels it joined while nobody was counting. When metering arrives, you want to argue from a usage number you already have, not discover it on an invoice.

Channel membership is now an access-control decision. An agent that reads the whole conversation reads the whole conversation: the vendor quote someone pasted, the customer's phone number in an escalation, the incident thread with the credential in it. Configure the role-based groups and model entitlements before you widen the channel list, not after someone asks what Claude has seen.

"When not to speak" is the feature that decides adoption. Most internal agent pilots die from noise, not from wrong answers. A four-way action choice that includes silence is the same design decision we make on every automation we ship: the default state is off, and speaking up has to earn its place.

Sources: Anthropic, VentureBeat.