On June 23 Anthropic shipped Claude Tag, and it isn't another model. It's a shift in how you work with AI at all. Claude now lives inside Slack: you tag @Claude in a channel the way you'd ping a colleague, and it takes the task and carries it to a result. No more "open a tab, re-explain the context, copy the answer back." Just a participant in your channel that already knows what's going on.
The number every outlet grabbed: 65% of Anthropic's own product-team code is now written by their internal version of Claude Tag. It sounds like a verdict on engineers. It isn't, and I'll explain why below.
At Gless we build AI agents for client tasks every week, so we read releases like this with one practical question: what here actually applies to a normal business that isn't on Claude Enterprise. Let's go through the facts, without the hype and without the panic. What Claude Tag is, what's genuinely new about it, what's reproducible right now, and where the catch is.
What Claude Tag is
Claude Tag is Claude embedded directly in Slack as a shared team member. The announcement landed on June 23, 2026. It's in beta for Claude Enterprise and Team customers and replaces the old "Claude in Slack" app: admins get 30 days to migrate, and eligible organizations receive an introductory launch credit.
Here's how it runs. An admin connects Claude to the workspace, grants access to specific tools and data sources, sets spend limits, and decides which channels the agent can operate in at all. After that, anyone in those channels tags @Claude in plain language: write this pull request, pull last month's sales numbers, run the analysis on this export. Claude breaks the task into stages, executes them with the tools it has, and comes back in the thread with a finished result.
Under the hood is Opus 4.8, the model Anthropic released less than a month ago. And the strongest evidence this isn't a demo: Anthropic now writes 65% of its product-team code with the internal version of Claude Tag. The same channels run their internal support and product analytics. The tool is proven on their own engineering, not just on a marketing page.
The real shift: AI as a colleague, not a browser tab
Technically, Claude could already do most of this. What's new is the format. AI stops being a separate place you visit and becomes a participant where the work already happens. Four things make the difference.
- Multiplayer. One Claude per channel, shared by everyone. Anyone can see what it's working on and pick up a thread where the last person left off. These aren't private chats where each person has their own bot with its own memory; it's one shared colleague with the channel's shared context.
- Memory and learning. Claude accumulates knowledge of the channel and the connected sources. You don't re-explain who you are, what the project is, and where the data lives on every request.
- Proactivity. In its "ambient" mode the agent surfaces relevant things on its own and circles back to stalled threads, instead of staying silent until directly summoned.
- Asynchronous work. You can queue a task hours or days ahead, and Claude runs it on schedule, with no one standing over it waiting.
In short: AI used to be a tool you picked up for a specific request. Here it's closer to an employee you delegate to and then forget about until the result lands. We wrote separately on why AI agents matter for your business, and Claude Tag shows the direction clearly: AI built into the process, not parked in the next tab.
What you can apply right now
One important caveat. Claude Tag is a beta for Enterprise and Team. No such subscription, no direct access yet. But the pattern the whole thing is built on is reproducible without it: an agent that lives inside a work tool and runs in the background 24/7.
Open agents like Hermes install on your own server, connect to messengers the same way, keep context between sessions, and run tasks on a schedule. We covered that in the post on AI agents that work 24/7 in messengers. The difference is maturity and who owns reliability, but the idea is exactly the one behind Claude Tag.
Who gets the most out of this format right now:
- Teams with a lot of repeatable routine living in the chat itself: digests, status updates, "pull me that number," "remind me what we decided on this account."
- People already living in Slack or Telegram who don't want to push colleagues into yet another separate interface for two actions a day.
- Small teams without a dedicated analyst or devops, where "ask the agent in the channel" closes a task in minutes instead of waiting for free hands.
The pattern is simple: the more predictable and repeatable the task, the faster the agent pays for itself. Rare, complex, judgment-heavy work is still wiser to keep with a human than to delegate into a channel.
Where the catch is: access, governance, and a human in the loop
Now honestly, because a polished release and a production process are different things. Anthropic didn't bake spend limits (org-wide and per-channel), separate access for different teams (sales and engineering see their own data and tools), and full audit logs of every action with the requester named into Claude Tag for nothing.
An agent with access to your data and tools, acting proactively in shared channels, raises immediate questions of security, privacy, and governance. Who sees what it writes? Which data does it touch while serving a harmless-looking request? You need to think this through before letting it into real processes, not patch it after.
About that 65% figure specifically. It's code the agent wrote, not code that shipped to production without a single human glance. Review, tests, and accountability for failures still sit with engineers. We covered why AI doesn't replace engineers in detail: Claude Tag confirms it rather than refuting it. It removes the routine, not the responsibility for the result.
Where's the line? Internal processes with a low cost of error are great candidates to hand to an agent in a channel and stop thinking about. Anything tied to money, customers, and production systems like CRM and payments needs reliability, integrations, and tests. That's engineering work on top of a convenient interface, not a replacement for it.
Where to start
You don't "roll out AI across the whole company" in one move. Take one channel and one repeatable task, give the agent narrow access scoped exactly to it, and watch for a week. Saves time with no surprises, expand to neighboring tasks. Doesn't, and you've spent one channel and a week, not a quarter and a budget.
If you want to figure out which of your processes you can genuinely hand to an agent like this, and which are better built as a reliable system with integrations and tests, that's exactly our AI implementation services.
FAQ
What is Claude Tag?
It's Claude embedded in Slack as a shared team member. You tag @Claude in a channel, it breaks the task into stages, executes them with the tools it has, and replies in the thread. Announced on June 23, 2026, running on the Opus 4.8 model.
Do you need Claude Enterprise to try it?
At launch, yes. Claude Tag is in beta only for Claude Enterprise and Team customers and replaces the old "Claude in Slack" app. But a similar setup, where an AI agent lives inside a messenger, can be built on open tools without that subscription.
Is it safe to let an agent like this into work channels?
With careful setup, yes. Claude Tag has per-channel access scoping, spend limits, and audit logs of every action. The core rule: give the agent the minimum access it needs and keep a human in the loop on critical tasks.
Will this replace employees?
No. Even at Anthropic, where the agent writes 65% of product-team code, the engineers are still there: on review, tests, and accountability for the result. An AI teammate removes the routine, not the decisions or the responsibility for them.
How is Claude Tag different from a regular Slack chatbot?
A regular bot answers a message and forgets the context. Claude Tag is shared across the whole channel, builds memory, works asynchronously on a schedule, and circles back to open questions on its own. It feels closer to a colleague than to a question-and-answer bot.
If you'd like to scope how to embed an agent like this into your own processes, get in touch and we'll look at your case.