Claude just moved into your Slack

Claude launched a feature that looks incredibly useful. Claude Tag.

You tag Claude inside Slack.

You let it follow the thread.

You connect it to HubSpot, Gong, Linear, project management tools, docs, and the systems your team already uses.

You ask it to pull data, update tickets, summarize decisions, schedule exports, and keep work moving while everyone else is doing ten other things.

Not anymore.

The old AI workflow was single-player. You opened a chat window, pasted context, got an answer, and copied the output somewhere else.

Claude Tags is the start of something much bigger: AI companies routing, remembering, and executing your work.

A multiplayer level.

The Trojan horse is the workspace, not the chatbot

Claude Tags is valuable because it stops making AI feel like a separate destination.

That matters.

If Claude lives inside Slack, it can watch the conversation where the work is happening. It can see the product thread, the sales handoff, the launch debate, the bug report, and the customer escalation.

In the demo, Claude can pull from HubSpot, Gong, and Linear. It can read and write. It can update tasks. It can schedule cron jobs. It can behave less like a chatbot and more like a teammate that never goes offline.

That is why the reaction was so big. The post had 15.4 million views because everyone can feel where this is going.

The interface is Slack. The product is memory.

The dangerous part is usage-based lock-in

This is a great product move from Anthropic.

It is also a dangerous bargain for enterprises.

The first layer is pricing. This is not the old per-seat SaaS model where you know what every employee costs. Once agents start pulling data, reading threads, writing updates, generating summaries, and running tasks in the background, usage compounds.

A teammate asks one question. Claude pulls five sources. It writes three updates. It monitors the thread. It follows up later. That single request becomes a chain of token usage across your company systems.

The second layer is much more important.

Knowledge lock-in.

Once an AI system sits in your channels long enough, it becomes the queryable repository for tacit knowledge. It knows why a launch date moved. It knows which customer objections keep showing up. It knows which internal debates already happened. It knows who approved what.

Agents can be copied. Models can be swapped. Company memory is much harder to move.

That is the piece operators need to pay attention to.

Your company brain needs its own infrastructure

I am not saying companies should avoid this direction.

The opposite.

This is where work is going. AI should live where the work happens. Slack, Teams, CRM, call recordings, docs, project management, support tickets, and internal decisions all need to connect.

But the memory layer should not casually belong to one model provider.

This is why we built Single Brain the way we did. It lives inside Slack and Microsoft Teams, but the goal is not to trap everything in one model. The goal is to build a company brain that compounds while still letting you route to the right model over time.

If a better open-source model comes out, you should be able to test it. If another model is better for a certain workflow, you should be able to route tokens there. If compliance matters, you should be able to host the infrastructure in a way that works for your company.

The brain is the asset. The model is one execution layer.

The operator play is to define the memory before the agents

Most teams will start by asking, “What can this agent do?”

That’s backwards.

Start with what the system should remember.

For example, when we work with Single Brain clients, we can ask it to look across the last 30 days of Gong calls, Granola notes, and HubSpot records to find customer sentiment. Then we can ask for exact examples that prove why clients are happy, unhappy, blocked, or asking for a specific thing.

That is not a generic summary.

That is operational memory turning into a management system.

The setup is simple:

1. Define the systems of record.
2. Define what the AI can read and write.
3. Define what needs to be remembered.
4. Define who has permission to ask for what.
5. Define which actions can run without human approval.

Do that before you let agents run loose.

Otherwise, you are just giving a fast worker bad context.

The sharpest lesson from Claude Tags

Claude Tags is not just a new AI feature.

It is a preview of how company operating systems will work.

The winning companies will not be the ones with the most AI tools. They will be the ones with the cleanest memory layer, the clearest permissions, and the best routing between humans, agents, and models.

Your AI stack will only be as good as the memory it compounds on.

Watch the full breakdown here: https://www.youtube.com/watch?v=8fjb1PHEd34 

To building your company brain,

Eric Siu