Claude Tag opens Slack beta for Team and Enterprise with memory and proactive follow-ups

• Anthropic opened ClaudeDevs' launch post and Boris Cherny's availability note say Claude Tag is a Slack beta for Claude Team and Enterprise, built as a persistent channel-level assistant with memory, tools, and proactive follow-ups.
• The internal version already writes a lot of Anthropic's product code: the ClaudeDevs thread says it now produces 65% of the product team's code, while Cat Wu's post says it merges 65% of product PRs.
• Anthropic is pitching Claude Tag less like chat and more like background operations: the launch thread lists incident response, bug triage, blocked work, postmortems, CI watching, and experiment monitoring as default patterns.
• Boris Cherny's sandbox explainer says each Slack thread gets its own isolated instance that can clone repos, write code, test, and compile, with separate memory and permissions per channel.
• Early hands-on posts from Thariq's workflow thread and Andrej Karpathy's reply frame the shift as organizational, not cosmetic: one Claude identity lives in the channel, keeps context, and works asynchronously alongside the team.

You can configure the agent's identity, open a Configure panel inside the Slack thread for connections and channel prompting, and even split personal and shared behavior because Noah Zwebner's reply says DMs act like your private assistant while channel Claude is its own teammate. Anthropic also showed six internal starter flows in Cat Wu's customization thread, and Noah Zwebner's GitLab note says GitLab support is already part of the rollout.

Anthropic's core design choice is simple: Claude belongs to the channel, not to the person who invoked it. Boris Cherny's launch thread calls it proactive, multiplayer, and equipped with its own identity and memory, while Noah Zwebner's launch post adds self-scheduled follow-ups and a richer tool set.

That changes the product from request-response chat into something closer to a standing teammate. Noah Zwebner's ambient-mode reply says Claude can pick work up at the channel level without being tagged every time, which is the feature that makes the whole thing feel less like a bot mention and more like delegated work.

Anthropic's own launch thread turned the use cases into a clean operating manual:

1. Incident response: ClaudeDevs' incident demo shows Claude pulling graphs, diffing the deploy, naming a likely root cause, opening the fix, watching recovery, and resolving the page.
2. Bug triage: the bug triage demo shows it picking up user reports, tracing the code path, reproducing the issue, writing a fix, and tagging the owner before merge.
3. Dependent work: the blocked-work example describes handing Claude tasks that must wait for another ship, then letting it return days later with an adjusted PR.
4. Postmortems: the postmortem example has Claude rebuild the incident timeline, draft the write-up in docs, and file action items as issues.
5. Background watchers: the CI watcher example uses threshold-based monitoring so Claude stays quiet until CI has been red too long, then surfaces the failing test and culprit commit.
6. Metrics and launches: the A/B test example gives Claude a metric and guardrails, then has it flag guardrail movement and prepare the rollout PR when results turn significant.

The useful detail here is how little of this is framed as prompt craft. Anthropic is selling recurring ownership, not better single-shot answers.

The execution model is stricter than the playful Slack surface suggests. Boris Cherny's sandbox explainer says every thread spins up its own isolated environment, where Claude can clone repos, write code, run tests, and compile, then discard the sandbox when the work is done.

That same post says memory and permissions are also scoped per channel, not globally. Boris Cherny's security outline breaks the safeguards into four layers: model training, classifier and auto-mode controls, limits on sites and credentials, and channel or workspace boundaries.

This is where the product stops looking like Slack glue. The architecture is really Claude Code, threaded through collaboration, with isolation and access control doing most of the heavy lifting.

Anthropic exposed identity as something teams can shape. Noah Zwebner's configuration post points to an identity setup flow for name, description, and system prompt, and the Configure-panel walkthrough says channel admins can edit prompting and connections from inside the thread.

The most concrete outside workflow notes came from Thariq's thread of favorites. He describes introducing Claude to each channel with a pinned message, treating that memory like a channel-specific Claude.md, creating a personal channel for forwarded work, keeping a pinned status message, and using emoji reactions such as ⏲️, ✅, ❓, and 🛑 to expose state at a glance.

Those habits matter because they show where the interface actually lives. A lot of the product is not in a settings page. It is in pinned instructions, thread rituals, and persistent memory.

Anthropic is not soft-pedaling internal reliance. the ClaudeDevs thread says the internal version writes 65% of the product team's code, including most of what built Claude Tag itself, and Cat Wu's post says it merges 65% of product PRs.

Andrej Karpathy, former Tesla AI director and OpenAI founding member, called this a third major LLM UI shift in Karpathy's post: first the model as a website, then as a desktop app, now as a persistent asynchronous entity with organizational tools and context. In his follow-up reply, he argues the hard part is making that arrangement actually work across tools, integrations, compute environments, memory, and security.

That is a strong claim, but Anthropic has at least chosen the right proof point. Shipping a channel-native agent is one thing. Claiming your own team already works from it all day is the bolder story.

Anthropic is already drawing a line between personal assistance and shared channel labor. Noah Zwebner's DM versus channel reply says a DM with Claude acts as your personal assistant with your own connectors and information, while messaging Claude in a channel invokes a separate shared identity that needs its own provisioned connections.

A few smaller rollout details also surfaced in replies. Noah Zwebner's GitLab note says GitLab companies are supported, Boris Cherny's reply says more surfaces are coming soon, and Boris Cherny's commit-attribution reply says commit attribution can be changed by asking Claude to update that setting.

Those details make the beta feel less like a single Slack integration and more like a permissions and identity system that happens to debut in Slack first.