There are two ways to add AI to a developer tool. Build the tool around the AI from the beginning, or bolt the AI onto a tool that was designed for a human writing code in a text editor and committing it to a version control system designed by Linus Torvalds in 2005. The first approach is winning. The second is producing demos that impress at conferences and die in production.
Cursor is the cleanest example of the AI-native approach. The editor is not a text editor with an autocomplete plugin. It is an AI agent operating environment. The AI has access to the entire codebase — every file, every function signature, every import path — and uses that context to generate code that passes type checking and matches the existing codebase's conventions. The interface is not a chat sidebar appended to an editor. The interface is the editor itself, with inline suggestions, tab-to-accept, and multi-file edits that span the project. The AI does not help the developer write code. The AI writes code. The developer reviews, edits, and accepts.
Claude Code, Anthropic's agentic coding tool, pushes the model further. The developer types a natural language instruction — "add rate limiting to the API endpoints using express-rate-limit, with different limits for authenticated and unauthenticated users" — and Claude Code opens the relevant files, writes the middleware, updates the routes, adds the configuration, writes the tests, and runs the test suite. The developer reviews the diff and approves. The workflow is not "code with AI assistance." The workflow is "specify the change, review the implementation, ship."
The legacy approach looks different and works worse. GitLab's AI features — code suggestions, vulnerability explanations, merge request summaries — are appended to the existing GitLab workflow of branches, merge requests, and CI pipelines. The AI can explain a vulnerability. It cannot prevent the vulnerability from being written in the first place because it is not integrated into the editor where the code is written. The AI can summarize a merge request. It cannot write the code that the merge request contains. The product's architecture — the core assumption that code is written by humans and managed through a branch-and-merge model — limits what the AI can do.
This is not a criticism of GitLab's engineering. It is a constraint of the product's DNA. GitLab was built for a world where humans wrote code in editors, stored it in Git repositories, and coordinated through merge requests. The AI features improve that world without replacing it. The AI-native tools replace the world. They assume that most code will be generated by AI, that the developer's primary role shifts from writing to reviewing, and that the tool's job is to make the AI as productive as possible rather than to make the human as productive as possible while the human does the writing.
The adoption data supports the AI-native thesis. Cursor reported single-digit millions of monthly active users in its mid-2026 update. GitHub Copilot reported 3.5 million paid subscribers. Claude Code, in a June 2026 usage report from Anthropic, was used by roughly 40% of the Fortune 500 for some class of development work. The legacy tools are growing AI feature adoption too — GitLab's AI features crossed 500,000 monthly active users in fiscal 2026. The gap in absolute scale is one thing. The gap in the depth of usage is another. A developer using Cursor or Claude Code generates a significant fraction of their daily code output from AI. A developer using GitLab's AI features uses the AI to review code they already wrote and to summarize things they already read. The value per user is orders of magnitude apart.
The implication for the developer tools market: the value is migrating from the collaboration layer (GitHub PRs, GitLab MRs, Jira tickets, Jenkins pipelines) to the creation layer (CDEs like Cursor, agentic tools like Claude Code, model providers that power them). The collaboration layer is not going away. It is becoming commoditized. The creation layer is the new value center. The companies that own the creation layer — or the AI infrastructure that powers it — will capture the economics. The companies that own collaboration and bolt AI onto it will maintain their existing revenue bases and lose the growth.
The legacy tools are not doomed. They are huge businesses with sticky customer bases. GitLab generated roughly $750 million in revenue in fiscal 2026, growing 28%. That is a great business. It is not the business that will capture the value created by AI-native development. Some other business will. It might be a model provider. It might be a new platform nobody has built yet. The market is sorting this out in real time. The sorting is faster than anyone expected.