Are you underestimating what Anthropic just launched?

This is a translation of a Norwegian article I initially published on LinkedIn, July 2, 2026.

Last week, Anthropic launched Claude Tag. In just a few days, of course, articles, blog posts and videos about the news have poured in. I had Claude and Firecrawl categorize the coverage for me. Much of it was about what Claude Tag is, and what it enables. No surprise there, because embracing this technology can bring substantial benefits.

Many articles were positive, some were skeptical, some were negative. But less than 3% of the coverage dealt with the thing that is concerning to me: lock-in. Not just because of costs, but because of the risk of losing the strategic freedom to choose your technology.

The memory layer is what matters

In my work building agentic systems, it has become very clear to me that the framework around the models will be what matters most in the time ahead. Not whether Claude or GPT or Gemini happens to be best at any given moment.

And at the center of that framework is the memory layer.
Where the ownerwhip of that lies will be decisive for your company's future.

When Anthropic accidentally published the Claude Code source code in April, there were some things discovered in the code that I found concerning, and so I wrote a blog post on how I saw those things as a foreshadowing of the direction we could expect Anthropic taking: they want to own as much as possible of the memory layer in our systems. When I wrote it, that claim was inferred from what had leaked.

Claude Tag is the confirmation.

A couple of weeks later, I went a bit deeper into what this memory layer is, and why I believe memory will be the central factor in agentic systems. I hinted then that Anthropic and its competitors "will make it easy for you to choose wrong." Just as Google has done for years, the AI companies will offer systems so convenient to use that choosing them is tempting. And building your own framework and memory layer isn't easy, and is of course more expensive in the short term than choosing Claude Tag.

The temptation of something that just works

What is Claude Tag?

Andrej Karpathy, who recently joined Anthropic, describes Claude Tag like this:

a self-contained, persistent, asynchronous entity with org-wide tools and context, working alongside teams of humans

Anthropic describes Claude Tag as a product that lets Claude participate in Slack "like a regular team member." Claude is invited to a channel, and then anyone on the team can @-tag it with a task. "@Claude look at this report and give me three ways to improve it." Or "@Claude based on the feedback in this channel over the past week, what 3 things should we focus on next week to reduce deviations in production?"

This isn't a chat you open to ask Claude to do something. Claude is where the work happens, follows the conversations and can take the initiative on tasks or suggest actions itself. And the central point is that Anthropic is building Claude's memory so it continually learns more about the company and how people work.

Tag also offers "Ambient mode," which means Claude reads and catalogs everything it can access, all the time. Not just when you tag it. Always. And the more Claude learns, the less you have to explain when you want a task done. It already has the necessary context as part of its memory.

Of course that's tempting. And of course it's tempting not to have to build this yourself. Why not just use what the experts have already built?

The battle for the workspace

That question requires looking beyond the individual solutions the AI companies provide and noticing the patterns behind them.

The industry is moving in this direction: in the future, the real value lies not in the models themselves, but in the systems around them. The harness. Anthropic and their competitors must take this path to avoid ending up as mere model providers. If you only supply the model, your only value proposition will be the model's quality. Once providers achieve truly self-reinforcing improvement, the differences between them will keep shrinking. And then the only remaining value proposition will be price. That's a downward spiral none of them wants to be in.

So they're working hard to build great harnesses around the models. Claude is now available as Tag in Slack. Cowork lets you use Claude for tasks in your own workspace, where your work already is. Claude desktop has connections to programs you already use. In practice, this means they're building more than just an integration here and there: they are building a coherent layer between Claude and the rest of the employees' work day.

Even though the models get most of the attention, models are no longer the primary thing they compete over. They're competing to own the workspace and the workflow. To become the place where the business thinks, collaborates, decides and innovates.

And every new feature or product you start using, increases the pain of leaving.

Lock-in: What it actually means

This is not just about technology, but every bit as much about strategic leeway.

You are no longer choosing an LLM model, but an operating system for knowledge work.

And if your company chooses that operating system, your employees' workflows, memory, automations, Slack threads, habits, projects, agents and identity will be built around one specific provider.

If you then want to change providers, it's not a matter of changing software, but of changing how people work.

Switching from Jira to Linear is painful, but it's still a task management system. Switching from AWS to Azure is harder, but both are platforms that offer much the same things.

Switching from Claude Tag to another AI ecosystem means changing how employees think, collaborate and delegate work. The company's specific version of "The Claude way of working." Changing providers then means retraining the whole business, but using worse tools than the ones you left. Because you're now starting from scratch.

Why from scratch?

Because the company's institutional memory has become proprietary.

Everything built into Claude's memory layer stays with Claude. Claude memory, Claude projects, Claude tasks, Claude workflow, Claude permissions, Claude history. Can you export all of this? What incentive does Anthropic have to offer that functionality? With a new provider, you therefore start from scratch and have to rebuild the whole intelligence layer. While developing new working habits at the same time.

Giving up ownership of the intelligence and memory layer is to make yourself so dependent on the provider that switching later is no longer a real option. It has gone from being a technical project to an organizational change.

That is why responsibility for this rests not just with the CTO, but with the company's entire management.

Price, tokens and freedom to choose

Little of the coverage I analyzed addressed lock-in. Of what I found, most was about costs and token usage. That's not the most important angle, but it is important. So let's talk about that too.

What happens if Anthropic decides to double token prices, and the business is deep in Claude-land? Your bargaining chip disappears. Today, most businesses can switch providers or models relatively freely. Once a business has become an all-out citizen of Claude-land, it has no bargaining position. Anthropic is free to dictate prices because migration is now an organizational problem more than a technical one. All you can do then is pay the price, and perhaps explaining to the board why.

And what would it cost if you do give Claude Tag full access to everything? How many tokens will be burned processing that much information continuously? Yes, Anthropic says it will offer features to limit token usage (token limits). But if "ambient mode" is to make sense, Claude Tag can't stop doing its job two weeks into the month because the budget has run out. Then everyone's work suddenly grinds to a halt.

If the business owns this layer itself, it can fine-tune the system freely. If Anthropic owns it, the business is at the mercy of the controls Anthropic chooses to provide.

Pace of innovation and model specialization

Today you can build an orchestration layer that lets you choose the model you prefer for each task, for example Opus for planning, GLM 5.2 for implementation, GPT 5.5 for code review, local models for confidential work, specialized models for translation, and so on. Claude Tag and similar solutions route everything through one provider, and the business loses the ability to optimize.

The business's pace of innovation becomes strongly tied to Anthropic's roadmap.

Innovation becomes dependent on the provider's pace of development. If another model provider makes an advance that is critical to your business, can you actually use it? Or do you have to wait and hope Anthropic offers the same thing?

Regulatory risk

What happens if regulatory requirements change to demand on-premise inference, EU-based models only, or national AI sovereignty? Can you migrate?

These are not hypothetical scenarios. Regulatory pressure in this direction is already well underway, and businesses that have given up ownership of the memory layer won't have good answers when those requirements arrive. And the events of the past month involving Anthropic's Fable and OpenAI's GPT 5.6 clearly show that government regulation is a real issue in several areas of AI.

These are new problems for everyone, and the biggest difference from before is the pace of change. Keeping up with developments is very challenging for most people, and impossible for those who don't work in AI. And FOMO is stronger than ever.

That's exactly why this is now less about making the best choices than about avoiding the worst ones - the ones that are hardest to reverse.

So what should you do?

I've used an awful lot of words to explain what you shouldn't do, and why. And to say that the alternatives are difficult and expensive. "Expensive" is relative, of course. It's about investing in systems that give you the benefits and value of adopting AI while letting you keep both control and ownership of your own data. None of this is new. You can absolutely build what Claude Tag and similar systems offer yourself. Or use (and contribute to!) open source initiatives trying to solve similar problems. The effect won't be as immediate as adopting ready-made systems from Anthropic, but you keep both ownership and the family silver.

Anthropic isn't launching Claude Tag because they're evil, but because it's the natural continuation of the path the industry has taken. Claude Tag isn't the problem either, but it's the clearest indication we've had yet of where the industry is heading. Anthropic's competitors are working toward the same goal, and we'll see many more launches like this.

So the question isn't really whether the AI companies are building walled gardens, but whether we are willingly walking into them.