title: "Claude Got Its Own Badge: It's Not Polite — It Simply Can't See" date: "2026-07-01" excerpt: '65% of Anthropic's product code commits are initiated by Claude. It has its own accounts, permissions, and memory boundaries. This is not just a tool update — it is a new signal for how AI agents enter organizational structures.' category: "ai-insights" showOnHomepage: true locale: "en" author: "Lincoln Wang" image: "/images/news/claude-tag-agent-permissions-2026.jpg" sourceVideo: "https://www.youtube.com/watch?v=VojDzHaciKQ"
Author: Lincoln Wang | Founder of MindsLeap | Global Partner at Founders Space | Founder of Founders AI Club
"Not because it's polite — because it simply can't see."
This came from Lidia on Anthropic's Cloud Code team. While introducing a new feature called Claude Tag, she described something that sounds small but carries enormous weight.
Someone pulled Claude into a legal channel and asked it to modify code. Claude didn't refuse — it couldn't. Because in that channel, it had zero access to the codebase.
This is not a prompt-level constraint. It is physical isolation at the system level.
65% of code commits — no longer initiated by humans
Lidia revealed a striking number during the demo: inside Anthropic, Claude Tag has already opened 65% of product pull requests.
In other words, the majority of code changes at this company are first set in motion by an AI agent.
She said: "For most of this year, this is how we've been pulling Claude in to do the work. Honestly, this is just how we work now."
Notice the cadence of that statement — not "we built a new feature," but "this is just what we look like now." An AI agent is no longer a tool you call. It has become the daily infrastructure of team collaboration.
A request that should never have needed a human relay
Lidia walked through a specific scenario worth unpacking.
A team was one week from launch. Engineer Drew learned from a sales rep that a major client was waiting on a feature: scheduled exports. Drew didn't own that module, so he looped in the owner.
Then he pulled Claude into the conversation.
Claude figured out on its own what scheduled exports meant and where in the codebase to make changes. It followed the group discussion, reacted to real-time product decisions, opened a pull request, and merged the changes into main.
Critically, Claude understood this change would affect the go-to-market timeline, so it automatically pulled the relevant teams into the sync.
There was no typical cross-department communication delay. Priya — the module owner — didn't even open Google Drive. Claude did the editing itself.
Claude got its own badge
This is where the story gets interesting.
Claude is not simulating human behavior. It has its own accounts, its own credentials, its own permission system.
Lidia kept repeating one line: "Claude has its own account. So every credential it uses is locked down."
What does this mean? Claude doesn't join a team as "some employee's assistant tool." It joins as an independent participant. It sees contracts in the legal channel and can edit code in the engineering channel. Pull it into the legal channel and ask it to modify code — it won't say "sorry, that's not appropriate." It literally cannot see where the codebase is.
Memory works the same way. What it learns in private channels or DMs does not leak elsewhere.
Permissions are not a moral problem — they are an architecture problem
This is the biggest takeaway from reading through this demo.
Over the past six months, the biggest concern for enterprises discussing AI agent deployment has always been: will it see things it shouldn't? Will it bring confidential information to the wrong place?
The mainstream answer: write better prompts to tell it not to cross boundaries; add a review layer so someone checks its output; use logs to monitor its behavior.
These approaches are essentially betting on the AI agent's "compliance." Anthropic took a completely different path — it built permissions into the architecture, not into the instructions.
"Not because it's polite — because it simply can't see."
Every enterprise leader considering AI agents should write this on a whiteboard. Because once you understand this, you realize: the real question was never "will AI obey?" — it is "have we designed the right organizational boundaries for it?"
When the AI agent becomes the N+1 person on the team
What Claude Tag reveals is not a smarter chatbot — it is a new collaboration topology.
Previously, the enterprise AI path was: employee hits a problem → opens AI tool → types a question → gets an answer → employee executes. AI always lived outside the employee's personal workspace.
The new path: someone in the group raises a question → someone pulls Claude in → Claude follows the conversation → it decides what to do, where to do it, and who to tell when done. Claude is inside the team's workflow, not outside it.
The difference between these two paths is not an efficiency issue. It is an organizational issue.
When an AI agent has its own identity, its own permissions, and its own memory boundaries, it is no longer someone's personal assistant. It becomes a node in the organizational structure — just like any other team member, with its own field of view, its own capabilities, and its own limitations.
The map is just starting to unfold
The most noteworthy aspect of Anthropic's demo is not what Claude can do — it is how it entered the organization.
Its own identity. Its own permissions. Its own memory boundaries. These three elements together constitute an entirely new paradigm for AI agent integration.
But this is just the beginning. The 65% figure comes from Anthropic's own team — a group that knows the product best, working in the environment most suited to it. Can it be replicated in a traditional manufacturer's supply chain coordination? In a retail company's store operations?
The answer is not yet clear. But it reads more as a signal: when AI agents start getting "badges," enterprises need to rethink not just which tools to adopt, but how to redesign their permission architecture, information flows, and collaboration boundaries.
The enterprises that earliest treat AI agents as "the N+1 person on the team" in their organizational design will be the first to make this work.
About MindsLeap
MindsLeap is an AI-native organization transformation accelerator.
In deep partnership with Silicon Valley innovation incubator Founders Space, MindsLeap connects cutting-edge global AI insights, the Silicon Valley tech entrepreneurship ecosystem, and real-world transformation scenarios for Chinese entrepreneurs.
MindsLeap is building a transformation ecosystem for entrepreneurs, startup founders, AI engineers, industry experts, and investors — helping enterprises move AI from awareness, strategy, and tools into organizational capability, business processes, product innovation, and growth systems.
This article was translated and adapted from the Chinese original with AI assistance.