|
ResearchAudio.io
Claude Just Became a Slack Teammate
One shared agent per channel, with memory scoped like permissions.
|
|
65%
Of product team code, per Anthropic
|
1
Shared Claude per channel
|
4.8
Opus version it runs on
|
|
The channel is the new terminal
|
|
Anthropic's product team gets 65% of its code from an agent that lives in Slack. Not a code editor, not a chat window: a channel. That figure comes from their own June 23 announcement of Claude Tag, and it refers to their internal version of the product.
|
|
The launch itself is simple to describe. Claude joins your Slack workspace as a team member, admins grant it access to specific channels plus whatever tools, data, and codebases they choose, and anyone in the channel can tag @Claude to delegate work. It breaks the task into stages, works through them, and replies in a thread. Anthropic calls it the beginning of an evolution of Claude Code, and the beta is live for Enterprise and Team customers.
|
|
I expected a coding pitch. The interesting half of the announcement is permissions.
|
Channel memory
Scoped to this channel
Admin-granted tools
Code, data, connectors
|
⇄ |
@Claude
One per channel, on Opus 4.8
plan → stage → thread
|
→ |
Thread replies
Visible to the whole channel
Self-scheduled tasks
Runs for hours or days
|
|
Source: Anthropic, Introducing Claude Tag (2026)
|
|
|
Four properties separate this from the bot you are picturing. It is multiplayer: within a channel there is one Claude that interacts with everyone, so anyone can see what it is working on and pick up where the last person left off. It learns: it builds context from the channels it sits in, and with permission it can pull from other channels and data sources, though Anthropic says it does not report from private channels.
|
|
It takes initiative: with ambient behavior enabled, it flags relevant information across its channels and tools, and follows up on threads that went quiet without resolution. And it is asynchronous: it can schedule tasks for itself and pursue a project over hours or days while you do something else. Anthropic says its own people now spend much more of their time delegating to many Claudes in parallel, and the pattern has spread past engineering into product metrics, support tickets, and bug root-cause work.
|
|
Here's the part nobody's talking about: the access model. Admins define a separate Claude identity per use, so a Claude set up for one team's work will not pass memories to the one set up for engineering, and will not give engineers a path into that team's data or tools. Memory, identity, and access all stop at the same wall, drawn at the channel. Every action lands in a log with the name of whoever requested it, and token spend can be capped for the org and for individual channels.
|
|
The key insight: memory is the permission boundary. Anthropic did not bolt access control onto an agent. It made the channel the unit of identity, recall, and data access at once.
|
|
|
|
|
DMs work too. Message Claude directly and it replies privately, using the personal tools and connectors you set up yourself.
|
|
It replaces Claude in Slack. Admins can opt in to migrate within 30 days, and Anthropic is issuing a launch credit to eligible Enterprise and Team orgs.
|
|
Two spend dials. Admins cap monthly tokens at the org level and per channel, which turns each channel into its own budget line.
|
|
Where it goes next. Slack is the start because that is where much of Anthropic's own day-to-day work already happens, and the stated goal is to expand tagging @Claude to the other places teams work.
|
|
|
|
The 65% line deserves a second read. It counts code created by their internal version on one team, which is not the same as code shipped, and the internal build has had a long head start the public beta has not. I still think the number matters, because it is the first hard figure a lab has put on channel-based delegation.
|
|
My take: the thing that looks like a Slack bot is really an identity system. Anthropic is provisioning agents the way IT provisions employees, with scoped access, an audit trail, and a budget. That template will outlive Slack, and it is the piece worth copying whether or not you ever use this product.
|
|
|
|
On an Enterprise or Team plan: the setup is four steps. Pair Claude Tag with your Slack workspace, grant it tools, cap your org's monthly spend, and test it in a private channel before letting it loose.
|
|
Building your own multi-user agent: make the room the memory boundary, not the user and not the org. The test takes five minutes. Ask your agent in room B what it learned in room A, and if it answers, you have a leak, not a feature. The channel-scoping audit checklist from this issue is in the paid archive in copy-paste form.
|
|
|
|
Ambient mode means a colleague-shaped process reads your team's channel all day and speaks up when it decides something matters. Where is your line: the metrics channel, the incident channel, or nowhere near anything human? Reply with the one channel you would never let it into. I read every response.
|
|
The unit of agent deployment is shifting from the chat window to the channel: one shared Claude, channel-scoped memory, an audit log, and a spend cap, provisioned like a new hire.
|
|
|
If the channel is the new terminal, the org chart is about to grow some agents.
|
|
On deck: what happens to context quality when an agent reads three months of Slack backlog. The token math surprised me, and I will show the numbers.
|
|
ResearchAudio.io · AI research, minus the noise
Sources: Introducing Claude Tag · Agent identity and access model · Claude Tag docs
|
