跪拜 Guibai
← Back to the summary

Claude Tag Turns AI from a Personal Chatbot into a Team Member Embedded in Slack

Claude Tag Is Here: AI Colleagues Are Moving from Chat Windows into Enterprise Workflows

Recently, Anthropic released a product worth paying close attention to: Claude Tag.

At first glance, it looks like an upgrade to Claude entering Slack: @Claude in a team channel, and it can read the context, break down tasks, call tools, and send results back to the thread.

But if you only understand it as "Claude in Slack," you might be underestimating this update.

In my view, what's truly interesting about Claude Tag isn't that it adds another entry point, but that it represents a clear shift in the product form of AI Agents:

AI is no longer just a chat window you open alone; it's beginning to become a "team member" embedded in an organization's collaboration workflow.

This has strong implications for enterprise AI, Agent Memory, knowledge bases, ticketing systems, R&D collaboration, and even future organizational management.

1. What Exactly Is Claude Tag?

Claude Tag is simple to use.

In a Slack channel or thread, team members can @Claude just like they would @ a colleague, and directly assign tasks:

This sounds like a chatbot, but it has a key difference from traditional chatbots:

Traditional AI assistants primarily work around "personal conversations," while Claude Tag works around "team context."

In a Slack channel, Claude is no longer just one person's private assistant, but a team-shared AI identity.

Zhang San asks it to analyze a problem, and Li Si can see the analysis process; Li Si continues to add context, and Wang Wu can also follow up and push forward. Claude's work process and results happen publicly within the team channel, not scattered in each person's own chat window.

This is a crucial point about Claude Tag:

It transforms AI from a "personal tool" into a "collaboration node."

2. This Isn't a Simple Upgrade of Claude Code

Anthropic officially views Claude Tag as part of Claude Code's evolution, because it can connect development needs from Slack directly to codebases, PRs, Issues, and engineering tasks.

For example, a team discusses a feature in a channel:

"We need to add a cadence picker to the product."

In the past, this requirement would typically go through a series of steps:

Product discussion in Slack → Engineer organizes requirements → Create a task in Jira or Linear → Open the codebase to analyze the scope of impact → Write code → Submit a PR → Return to Slack to sync progress

Claude Tag tries to compress this chain into:

@Claude in Slack → Claude reads the discussion context → Analyzes the codebase → Breaks down the task → Generates a plan or draft PR → Returns to the original thread to sync results

This is no longer just about "being able to write code"; it's about AI starting to enter the real R&D collaboration pipeline.

More importantly, Claude Tag isn't just for engineering teams.

Officially mentioned scenarios also include querying product data, handling support tickets, preparing for customer meetings, monitoring channels, and organizing action items. In other words, code is just one high-value scenario; the real target is the larger enterprise workflow.

So I prefer to understand Claude Tag as:

An enterprise-grade Agent with Slack as its entry point, organizational context as its foundation, and tool invocation as its execution capability.

3. Four Keywords: Shared Context, Persistent Memory, Proactive Intervention, Asynchronous Execution

What's most noteworthy about Claude Tag isn't "@ it and get an answer," but the four capabilities behind it.

3.1 Shared Context: AI Begins to "Read the Team Scene"

In the past, when using AI, we often had to provide a lot of background first:

"Our project is like this..." "The discussion just now was about this issue..." "Someone said something before..."

This is actually quite unnatural.

In real team collaboration, a lot of knowledge isn't in formal documents, but flows constantly within Slack, Feishu, WeCom, meeting minutes, PR comments, ticket discussions, and CRM records.

Claude Tag's first step is to let AI enter these collaboration scenes.

It can read the context of channels and threads, understand what everyone has discussed, who is responsible for what, which issues remain unresolved, and which decisions have been made.

This means AI no longer relies solely on user-input prompts, but can obtain context from the organizational collaboration process.

3.2 Persistent Memory: AI No Longer Starts from Scratch Every Time

Another key aspect of Claude Tag is that it accumulates team context over time.

For example, items mentioned in Monday's standup can still be remembered by Claude on Thursday; project background discussed in a channel last week doesn't need to be re-explained each time; the team's tech stack, business habits, and division of responsibilities can gradually become the basis for its understanding of work.

This is very close to the Agent Memory problem I've been focusing on.

The problem with many AI assistants in the past was: they could answer well, but every time it felt like their first day on the job.

To truly enter the enterprise scenario, AI can't always be like a "temporary outsourcer." It must gradually understand the organizational state:

So what's important about Claude Tag isn't just "remembering," but that it makes organizational memory a core capability of enterprise Agents.

3.3 Proactive Intervention: AI No Longer Waits for Questions

The interaction mode of traditional AI assistants is:

Person asks a question, AI answers.

Claude Tag has taken a step forward.

With the relevant mode enabled, it can proactively remind the team:

This is actually a significant change in the product form of Agents.

AI has transformed from a "passive responder" into a "proactive observer."

Of course, if this capability isn't done well, it could also become noise. So the key for future enterprise Agents isn't just "can it be proactive," but:

When should it be proactive, when should it not disturb, and when must it escalate to a human.

This requires very mature permission, priority, context judgment, and responsibility boundary design.

3.4 Asynchronous Execution: AI Begins to Undertake Long-Term Tasks

Claude Tag also has a very important point: asynchronous execution.

In the past, many AI tasks were synchronous. You ask a question, wait for it to generate a result, and then proceed to the next round.

But real work isn't like that.

In real work, many tasks span hours, days, or even longer:

Claude Tag's direction is to allow AI to undertake these long-term tasks within team collaboration.

This is more like a true Agent: it doesn't just complete a single answer, but can continuously advance towards a goal.

4. What's Really Being Fought Over Is "Organizational Knowledge"

Why did Anthropic create Claude Tag?

On the surface, it's about integrating Claude into Slack.

But looking deeper, it's competing for the most important and hardest-to-structure asset within an enterprise:

Organizational knowledge.

A lot of truly valuable information in enterprises isn't in formal documents.

It's hidden in:

This information typically has several characteristics:

First, it's scattered. Second, it's dynamically changing. Third, it carries strong context. Fourth, it's often not formally captured. Fifth, it's hard for both newcomers and AI to understand.

This is also why enterprise knowledge bases have always been difficult.

Many companies think a knowledge base is just putting documents into a vector database and connecting it with RAG.

But the reality is:

Enterprise knowledge is not a static collection of documents, but a continuously changing system of organizational state.

Claude Tag's direction is precisely to try to let Claude enter this state system.

This is also the direction that enterprise AI players like Microsoft Copilot, Glean, Databricks, and Snowflake are all competing for: whoever can understand enterprise context has a better chance of becoming the enterprise AI entry point.

5. From "People Find Systems" to "People Find AI, AI Finds Systems"

The old way of using enterprise software was:

I want to check a customer, go to CRM; I want to see tasks, go to Jira; I want to check code, go to GitHub; I want to see data, go to BI; I want to find documents, go to the knowledge base; I want to communicate, go to Slack or Feishu.

Each system has its own entry point, permissions, processes, and interface.

But the trend represented by Claude Tag is:

People no longer directly find systems; they find AI, and AI then calls systems based on the task.

This will bring a big change.

In the future, employees may not need to remember how to use dozens of systems, but only need to express their goal in the collaboration entry point:

"Help me check what risks this customer has recently."

Then AI automatically checks CRM, emails, meeting minutes, tickets, contracts, and historical communication records, finally compiling the results into a readable brief.

Or:

"Why does this online issue keep recurring?"

AI automatically looks at monitoring, logs, code changes, PRs, incident records, and recent alerts, organizing possible causes, impact scope, and suggested actions.

This is where enterprise AI Agents truly provide value:

It doesn't replace a single tool, but becomes a unified execution layer between multiple tools.

6. But What's Really Hard Isn't Connecting Tools, It's Governing Context

Claude Tag looks cool, but it also exposes the hardest set of problems for enterprise Agents.

6.1 Permission Governance

Which channels can AI see? Which tools can it access? Can it view sales data? Can it view engineering code? Can it read information across departments? Can it execute actions on behalf of users?

These are not simple technical problems; they are enterprise governance problems.

Anthropic introduced different Claude identity designs in Claude Tag: different channels, teams, and tasks can have different Claude identities and permission scopes.

This direction is correct.

Because once enterprise AI enters real workflows, it can no longer use the permission model of a personal chatbot.

It needs Agent Identity:

Otherwise, the stronger the enterprise Agent, the greater the risk.

6.2 Memory Governance

Claude Tag continuously accumulates context, but continuous accumulation itself doesn't equal correct understanding.

For example, if a project's responsible person changes, does AI know the old person is no longer responsible? If a requirement direction is overturned, will AI still reference the old decision? If a customer's status has changed, can AI still distinguish between historical information and current facts? If an incorrect conclusion appears in a channel, will AI remember it as organizational knowledge?

So the core of long-term enterprise Agents isn't just Memory, but Memory Governance.

This is what I've been emphasizing:

It's not about "remembering more," but about "correctly governing change."

Enterprise Agents need to know which information is active, which is historical, which has gone stale, which requires human confirmation, and which cannot be propagated across boundaries.

Otherwise, the more "memory" it has, the more likely it is to bring outdated information, incorrect information, or out-of-scope information into subsequent tasks.

6.3 Responsibility Boundaries

When Claude Tag creates a PR, modifies a ticket, or notifies a responsible person, who does the responsibility belong to?

Is it the initiator? The channel owner? The administrator? Or AI itself?

This issue will become increasingly important in the future.

Enterprise Agents are not ordinary chatbots; they connect to real systems and produce real actions. As long as they can act, there must be responsibility boundaries.

So I believe the maturity of an enterprise-grade Agent can be judged by a simple formula:

Agent Capability = Model Capability × Tool Capability × Context Capability × Governance Capability

If any one of these links is too weak, problems will arise.

Strong model but chaotic permissions leads to security issues. Many tools but poor context leads to doing the wrong thing. Long memory but poor governance leads to referencing outdated information. Strong proactivity but poor boundaries leads to noise or even accidents.

7. Inspiration for Domestic Enterprise AI Products

Claude Tag offers several very direct inspirations for teams in China working on AI Agents, intelligent customer service, knowledge bases, and enterprise process automation.

First, the entry point must return to the workplace.

What enterprise employees use most frequently isn't a new AI App, but WeChat, WeCom, Feishu, DingTalk, Slack, email, ticketing systems, and business backends.

For AI to create value, it can't stay forever in a separate chat window; it must enter real workflows.

Second, Agents should be designed around teams, not individuals.

Personal assistants solve "my" problems; enterprise Agents solve "our" problems.

This means they must support shared context, multi-person handoffs, visible processes, and traceable results.

Third, knowledge bases can't just be RAG.

A lot of enterprise knowledge isn't in documents, but in processes and collaboration.

Future enterprise knowledge bases must upgrade from "document retrieval systems" to "organizational state systems."

They shouldn't just answer "where was this mentioned," but also:

Fourth, governance capability will become a core competitive advantage for enterprise Agents.

Model capabilities will become stronger, tool invocation will become more standardized, and what truly differentiates players will likely be:

Enterprises don't lack a "better chatting" robot; they lack an AI collaborator that can work stably within complex processes.

8. My Judgment on Claude Tag

Claude Tag is not the end point, but it points to a very clear direction:

The competition in enterprise AI is shifting from "whose model is stronger" to "who better understands organizational workflows."

Claude Code made many developers feel the power of AI writing code.

What Claude Tag aims to do is go a step further:

Let Claude not just write code, but understand where requirements come from, where discussions happen, who participated in decisions, how tasks are broken down, and how results return to the team.

This is why it's truly worth paying attention to.

From chat window, to IDE; From IDE, to Slack; From personal assistant, to team collaborator; From one-time answers, to long-term asynchronous execution; From tool invocation, to organizational context understanding.

This line is very clear.

AI Agents are moving from "being able to do tasks" to "participating in organizations."

And once AI begins to participate in organizations, the real challenge is no longer model reasoning ability, but:

This is also why I think Claude Tag deserves a separate article.

It's not just a product update from Anthropic, but a signal:

The future AI colleague may not sit in a standalone App. It's more likely to appear where the team works every day, read the context, catch tasks, advance continuously, and then @ you again when it needs your decision.

The true era of enterprise Agents may have begun with this @Claude.

References:

[1] Anthropic: Claude Tag Product Page https://claude.com/product/tag [2] Anthropic: Introducing Claude Tag https://www.anthropic.com/news/introducing-claude-tag [3] Reuters: Anthropic launches Claude Tag in Slack with plans for wider rollout https://www.reuters.com/technology/anthropic-launches-claude-tag-research-preview-slack-users-2026-06-23/ [4] TechCrunch: Anthropic’s Claude Tag is learning your company, one Slack message at a time https://techcrunch.com/2026/06/23/anthropics-claude-tag-is-learning-your-company-one-slack-message-at-a-time/ [5] Slack Marketplace: Claude & Slack Integration https://slack.com/marketplace/A08SF47R6P4-claude