Claude Tag and the Third Era of Working With AI

For three years, working with a large language model has meant the same quiet ritual. You stop what you are doing, open a tab or an app, type your question into an empty box, wait, then copy the answer back into the work you were actually doing. On 23 June 2026, Anthropic released something built to end that ritual, and one of the most respected voices in AI says it marks the third time the whole shape of this technology has changed.

The product is called Claude Tag, and the idea is almost mundane until you sit with it. Anthropic introduced it as a way for Claude to join a team's chat as something close to a coworker. It starts on Slack: an administrator gives Claude access to chosen channels and connects it to the tools, data and even codebases the team uses. From then on, anyone can type @Claude, hand it a job in plain language, and get back to their own work while it gets on with the task, posting back in the thread as it goes. It is in beta for Claude Enterprise and Team customers, and it runs on Anthropic's Opus 4.8 model.

What makes this worth a considered look is not the feature list. It is that Anthropic is not really describing a feature at all. It is describing how the company now works. By its own account, tagging Claude has become one of the main ways things get done inside Anthropic, and 65% of its product team's code is now created by an in-house version of the tool. Staff tag it to chase down product metrics, work through support tickets, and hunt for the root cause of tricky bugs. When the company that built the model quietly reorganises its own days around it, that tells you more than any benchmark.

The third redesign of how we use AI

That is the framing Andrej Karpathy reached for too. Karpathy, a founding member of OpenAI and Tesla's former head of AI, is one of the few people whose read on where this technology is heading the rest of the industry actually waits for. Replying to the launch, he argued this is the third major redesign of how humans interact with language models, and the lineage he drew is the clearest way to understand why it matters. In the first era, the model was a website you went to. In the second, it became an app you downloaded onto your computer. This third one is different in kind, not degree.

The jump from the first era to the second was mostly about convenience: the same chatbot, closer to hand. The jump Karpathy is pointing at is about identity. The model stops being a thing you consult and starts being a participant. It has a presence in your workspace, a memory of what has happened there, and the ability to act over time rather than only answering when spoken to. That is a genuinely new relationship, and most of us have not built the instincts for it yet.

From a tab you open to a teammate you tag

Anthropic's own description of what is new lines up with the bigger claim. Within a channel there is one Claude that everyone shares, so a colleague can pick up a task where someone else left off, the way you would with a person rather than a private chat window. It learns over time, building context from the channel and the data sources it is allowed to see, so nobody has to brief it from scratch every morning. If its ambient behaviour is switched on, it takes initiative: flagging information it thinks you need and following up on threads that have gone quiet without being asked. And it works asynchronously, scheduling its own tasks and pursuing a project across hours or days while the humans get on with other things.

Stack those four traits together and you get the shift in plain terms. Shared, not private. Remembers, rather than forgets. Acts, rather than waits. Runs in the background, rather than only in the moment. None of those on its own is dramatic. Together they are the difference between a tool and a teammate.

What it means for a small business

Here is the honest caveat first, because it matters. Claude Tag is in beta and limited to the Enterprise and Team tiers today, so a solo operator or a small Australian shopfront cannot simply switch it on this week. But treating that as a reason to look away would be a mistake. These tiers are almost always the front of a queue rather than the end of one, and the underlying shift, AI that lives inside the tools you already use, is showing up across the whole industry at once, not just in one product.

The deeper point is about how you think, not which button you press. For a small team the bottleneck is rarely ideas, it is hands. When you are the owner, the salesperson, the bookkeeper and the support desk at once, an always-on colleague who can be handed a repeatable job and trusted to work through it is the kind of help you could never previously afford. The mental model has to move with the technology: stop picturing AI as a place you go, and start picturing it as someone in the room. Here is what that unlocks once it is set up properly:

• AI stops being a detour. The help arrives inside the conversation where the work already happens, instead of in a separate tab you have to remember to open.
• The context compounds. An assistant that follows along stops needing to be re-briefed every day, so the tenth task it does for you is better than the first.
• Work happens while you are doing something else. Jobs can be set running and picked up later, so a small team effectively gains hours it never had.
• One shared brain replaces ten private chats. The whole team works with the same assistant and the same context, instead of everyone quietly prompting their own.
• The scarce skill changes. The hard part becomes choosing what to hand over and judging what comes back, not the doing itself.

The shift is not that you can talk to an AI. It is that the AI now sits in the room with your team, remembers the conversation, and gets on with the work while you do something else.NextAura

This is also where the easy version of the story falls down. An assistant that shares your team's context and can act on its own is only as safe as the edges you give it: what it can touch, what needs a human to sign off, which part of the business it is allowed to see. Anthropic has clearly thought about this, letting administrators scope each Claude's tools, data and memory tightly per channel and keep a log of everything it has done. That is the right instinct, and it is also the part that decides whether one of these teammates saves you real hours or quietly creates a new mess. We went deeper on that question in our companion piece on what to hand an AI teammate first.

Karpathy's own verdict was that it takes a while to wrap your head around, but that it works and it is awesome. He is probably right on both counts. The companies that get the most from this will not be the ones who rushed to plug in a bot. They will be the ones who used this window to rethink which parts of their work could be handed to a tireless colleague, and how to keep that colleague pointed in the right direction. That is a strategy question long before it is a software one.

This is the shift we help Australian small businesses get ahead of. At NextAura we design and build the AI agents and ways of working that let a small team hand real work to AI without losing the thread: deciding what to delegate, wiring in the right context, scoping what it can reach, and keeping a person in charge of the judgement calls. If you would rather be ready for the colleague-in-the-room era than scrambling when it lands in your tools, read our companion piece on what to hand an AI teammate first, then get in touch and we will help you put it to work.