Claude Tag introduces agent identity for Slack channels and DMs
Anthropic detailed Claude Tag's agent identity model, giving Claude its own credentials in shared channels and user identity in DMs. Teams can share one Slack agent without mixing multiple people's permissions or histories.
TL;DR
- ClaudeDevs' agent identity thread says Claude now uses its own credentials in shared Slack channels, while Noah Zweben's reply says DMs still run on your personal connectors and identity.
- In channels, ClaudeDevs' walkthrough and Boris Cherny's sandbox note describe one shared Claude with per-channel permissions, per-thread sandboxes, and audit logs under Claude's own account.
- ClaudeDevs' launch thread framed Claude Tag as multiplayer, async, and proactive, and Noah Zweben's reply added that teams can set how ambient it should be without tagging it every time.
- Internal usage is part of the pitch: ClaudeDevs said Anthropic's product team used its internal version all year, while Cat Wu's post and Boris Cherny's post both put the share of product PRs at 65%.
You can read Anthropic's launch post, watch the agent identity setup video, and open the per-channel Configure flow. The interesting bit is the split: ClaudeDevs' thread treats DMs like a personal assistant and channels like a new employee account, while Thariq's best-practices thread shows teams already using pinned instructions, status emojis, and channel-specific memory like a lightweight operating manual.
Agent identity
The core product decision is simple: in a shared channel, Claude acts as itself.
According to ClaudeDevs' thread, that avoids the permission mess of borrowing one teammate's credentials in a thread with several humans. The same thread says DMs stay personal, with Claude using your own email, calendar, and connectors, while channels get Claude's own GitHub, Linear, or warehouse account.
That split is the whole form factor. Andrej Karpathy's post called it a third major LLM UI shift, and Greg Isenberg's post pushed the same idea further by describing agents as coworkers with their own logins, history, and accountability.
Permissions and memory
Anthropic's answer to the obvious security question is scoped access plus a single identity to audit.
ClaudeDevs' thread says admins provision Claude once, then limit what it can reach by channel. In the same thread, Anthropic says every action lands in the connected system's own logs under Claude's account, and revoking that account cuts access everywhere at once.
The memory model is also channel-shaped. Thariq's best-practices thread says each channel's Claude is different, with pinned instructions, persona setup, and remembered norms that behave like a shared Claude.md. Shann Holmberg's comparison describes the same pattern as persistent, scoped memory with no bleed across roles.
Proactive workflows
Anthropic is selling Claude Tag less as chat and more as background labor.
The launch thread breaks that into concrete jobs:
- Incident response: Claude pulls graphs, diffs deploys, proposes root cause, opens a fix, and watches recovery, per ClaudeDevs' incident example.
- Bug triage: it can sit in a feedback channel, reproduce issues, write a fix, and tag an owner, per ClaudeDevs' bug triage example.
- Dependent work: it can wait on a backend ship, then come back later with the frontend PR, per ClaudeDevs' dependent work example.
- Background watching: it can monitor CI or experiment guardrails and only post when a threshold trips, per ClaudeDevs' watcher examples.
Practitioner posts add a few operating patterns. Thariq's thread recommends a personal channel, a pinned running status message, and top-level status emojis. Allie K. Miller's example describes one Claude channel per teammate, where the agent auto-executes low-risk tasks and escalates higher-risk work with emoji approvals.
Configuration
The product already exposes more control than the launch post suggests.
Noah Zweben's follow-up says each channel has a Configure link for connections and channel prompting, and notes that admins can disable channel prompt edits. The same post says repository selection is still "coming soon," which is a useful limit on day one.
Two smaller rollout details showed up in replies. Noah Zweben's reply says teams can tune how proactive Claude should be, including ambient monitoring without repeated tagging, and Boris Cherny's launch thread says Slack is only the first surface, with more surfaces coming soon.