Video summary

Inside Anthropic: How Claude Tag Is Changing Agentic Work

Main summary

Key takeaways

Technology

Summary of the video (technological concepts + product features)

Claude Tag: what it is and why it’s different

  • Claude Tag is positioned as a “proactive teammate” that lives in Slack (not just something you manually trigger).
  • It includes the connectors/tools/context familiar from Claude Code, and adds:
    • Proactivity: can initiate work or return to you when attention is needed.
    • Persistence / long-running execution: can carry tasks through longer time horizons.
    • Collaboration/multiplayer behavior: can interact with multiple teammates at once, gradually building shared context.
    • Memory/context spanning channels: can maintain context at least at a channel-level scope, and (optionally, depending on permissions) consider other public channels.

How Claude Tag fits with Claude Code (Slack integration vs agent orchestration)

  • Anthropic uses Claude + Claude Code integrated into Slack as an existing pattern (“Claude in Slack”).
  • The speaker distinguishes Claude Tag from simply tagging Claude in Slack:
    • Claude Tag runs an end-to-end workflow more autonomously (including ticketing → coding → PR → review/ping).
    • Instead of you launching Claude Code, Claude Tag can spin up its own sandbox, run code in the repository (where permitted), and notify when a PR is ready.
    • It can also verify against requirements (e.g., derived from the ticket/issue) and prompt for reviews.

Example end-to-end workflow described (Slack → Linear/Tickets → PR)

A typical flow:

  1. Discussion begins in Slack (feature/bug/request).
  2. Claude Tag uses thread/context to propose or drive creation of a Linear ticket.
  3. It can assign/ping relevant teammates for coordination.
  4. It executes implementation using coding tools and outputs a pull request.
  5. A human approval/review can be inserted as a configurable gate (e.g., require explicit approval before merge/deployment).
  6. It can prompt reviewers and ensure alignment with requirements.

“Positioning system” + permissioned access (enterprise-ready design)

Claude Tag uses a positioning/memory scoping system:

  • Context is tunable, primarily scoped to a channel, with the ability (via permissions) to access additional public channels.

For scaling and safety, permissions are a core design principle:

  • Each channel/workspace has its own permission scopes (what tools/connectors/APIs it can access; what other channels it can read).
  • Agent identities: unlike Claude Code patterns where actions use your personal permissions, Claude Tag uses its own agent permissions/keys—improving auditability and making it more team-based than person-based.

Operational/behavioral shift: chat-based async agents vs IDE-based sessions

The speaker contrasts:

  • Claude Code as more “single-player” / session-based, where the developer is present at each turn.
  • Claude Tag as shifting toward:
    • less session-bound, more amorphous workflows
    • asynchronous long-running tasks (e.g., returning after hours with a completed feature)
    • a more multiplayer, collaborative working model in Slack

This is framed as a trust transition:

  • moving away from “seeing every code step” toward relying more on tests/validation.

Trust and capability trends (agent runtime doubling)

The speaker references research (METR) indicating:

  • every ~4 months, agents’ successful autonomous task time horizon roughly doubles
  • improvements come from both model capability and better ability to verify results at multiple stages

Impact metrics / internal adoption at Anthropic

  • After deploying Claude Tag internally: 65% of product teams’ PRs are opened by Claude Tag.
  • For Claude Code (earlier productization milestone): ~half the company used Claude Code weekly (internal “PMF” signal).

Internal use cases beyond engineering

Broader organization-wide uses include:

  • Marketing workflows (ad/copy generation pipeline)
  • Incident response triage and routing to the right people
  • Sales weekly briefs generated on a schedule
  • Team dashboards / reporting
  • Prototype feedback loops, where agents work “in the background”

Claude Tag proactivity can be tuned on a spectrum:

  • respond-only → scheduled → proactively jump into relevant threads.

“Dreaming” (managed agents research preview)

  • Dreaming is a research preview feature in Anthropic’s Managed Agents offering.
  • Goal: continual learning/optimization of an agent’s memory and performance over time.

Mechanism (as described):

  • provide memory stores + session transcripts (“traces”)
  • a separate agent reviews for:
    • discrepancies
    • missing/misleading context
    • suggestions for reorganization
  • it produces hypotheses, attachments to evidence, and lets you choose which changes to implement

Outcome:

  • tasks can run over time where agents get better day-to-day.

Practical “day-to-day” recommendations from the speaker

  • Set up Claude Tag and try it quickly (requires Slack admin + configuration).
  • Example starter configurations mentioned:
    • Daily Briefing: curated digest of the past 24 hours across international teams
    • Urgent live pings: only notify on matters relevant to you
    • Weekly team reports to reinforce shared context
    • optionally: have Tag build custom software/dashboards, deployed via Claude Code artifacts

Main speakers / sources

  • Primary source: Lamis (member of technical staff at Anthropic, Applied AI team)
  • Host: Simon Maple (AI Native Dev podcast)
  • Co-host mentioned / podcast hosts: Guy Podjarny (mentioned in intro)

Original video