Summary
Anthropic product manager Noah walks through the agent identity model powering Claude Tag, a proactive multi-user agent designed to collaborate across Slack teams. The video explains the core design problem: when Claude operates in a channel with multiple users, traditional per-user permission models break down — especially when Claude is synthesizing work from several teammates or acting on a monitoring alert with no human requester at all.
The solution is to give Claude its own identity: a dedicated agent account with credentials scoped specifically to what administrators want it to access. Permissions are structured in three tiers. A workspace-level access bundle sets the baseline — tools every workspace member is already allowed to use, provisioned once and inherited by every channel. Channel-level step-ups allow individual teams to grant access to more sensitive resources, such as a data warehouse, scoped only to that channel. DMs with Claude run under the user’s own personal credentials, making them effectively a private assistant context.
The walkthrough includes a live demonstration of the Claude Tag admin UI, showing how access bundles are created, credentials are tested, and bundles are attached to workspace or channel scopes. The model is designed to make Claude’s capabilities predictable for users and auditable for security teams, with access that never changes based on who happens to ask a question in a thread.
📺 Source: Claude · Published June 23, 2026
🏷️ Format: Keynote Launch
