Most AI at work still behaves like a chatbot in a side panel: you open it, you ask, you copy the answer back into wherever the work actually lives. Anthropic's newly announced Claude Tag points at a different model — and it is worth understanding regardless of which vendor you use.
The idea: Claude joins your team. Starting in Slack, you grant it access to selected channels and connect it to chosen tools, data, and codebases. Then anyone can tag @claude and hand off a task while they do other work. Anthropic says 65% of its product team's code is now created by an internal version of this.
What makes a "tagged agent" different from a chatbot
- Multiplayer — one agent per channel that everyone shares; anyone can see what it is doing and pick up where a colleague left off.
- Learns over time — it builds context from the channels and data it is permitted to see, so people stop re-explaining background on every request.
- Takes initiative — with "ambient" behaviour on, it proactively flags things and follows up on threads/tasks that went quiet.
- Works asynchronously — set it a task and it works while you focus elsewhere; it can schedule its own work over hours or days, many agents in parallel.
That is closer to a coworker than a tool.
The hard part was never the model — it's governance
An AI teammate with access to your channels, tools, and data is exactly as useful as it is risky if the access model is sloppy. The most important detail in the announcement isn't capability — it's control:
- Admins scope which tools and information the agent can use, per channel.
- Each setup is effectively a separate identity — a sales-configured agent never leaks memory or data into an engineering one.
- Token-spend limits per org and per channel.
- An audit log of everything the agent did and who requested it.
That list is the real checklist for deploying any agent in a regulated, data-sensitive business. Before you tag an agent into a channel touching customer PII, financials, or contracts, you need scoped least-privilege access, memory boundaries between contexts, full audit trails, spend caps, and human approval on irreversible actions.
The governance and orchestration layer — not the underlying model — is what turns an impressive demo into something you can run in production.
The takeaway
Claude Tag matters less as a single product and more as a marker of where applied AI is heading: agents that live inside your workflows, accumulate context, and act on their own initiative within tight guardrails. If you want an AI teammate in your operational channels, start with the access model, the audit trail, and the orchestration around it — not the demo.
Full version, with the real-estate / PropTech angle, on the VSBD blog. Source: Anthropic — Introducing Claude Tag.
Top comments (0)