Shopify CEO Threatens Claude Code Ban Over AGENTS.md

Shopify CEO Tobi Lütke just turned a Markdown file into a weapon.

On X, he said he is considering banning Claude Code inside Shopify unless Anthropic makes it read AGENTS.md and the .agents/skills directory. His complaint is not about model quality. It is about “split brain”: when engineers use Claude Code alongside Codex, Cursor, Gemini CLI or Copilot, each tool reads a different project rulebook. The same repository ends up running under two sets of instructions.

The Claude Code team, represented by Thariq Shihipar, quickly replied that they are making the tool “more customizable” and that support for AGENTS.md is coming. But the exchange exposed something larger. The coding-agent market is moving from a features race to a control-plane standards war, and the winner may not be the best model but the one that speaks the repository’s native rules.

What AGENTS.md actually is

AGENTS.md is best understood as a README written for AI employees. It sits at the root of a repo (or under specific directories) and tells coding agents how the project is built, how to run tests, which directories are off limits, what security checks must pass before a commit, and what skills are available. It is an attempt to encode an organization’s development policy into the codebase itself.

The format is gaining traction fast. OpenAI Codex, Cursor, Gemini CLI, GitHub Copilot, Devin and VS Code already support it. In December 2025, OpenAI donated AGENTS.md to the Agentic AI Foundation under the Linux Foundation, explicitly to keep it vendor-neutral. The foundation now governs the spec. Anthropic is a co-founder of that same foundation and donated its own MCP protocol there.

Claude Code, however, uses its own system: CLAUDE.md for project rules and .claude/skills for skills. Anthropic’s argument is that different model families behave differently under different system prompts. Claude has specific preferences for how skills, rules and context are organized, and Claude Code tailors prompts accordingly. The company calls this “context engineering.” In July, Shihipar noted that the team had already removed about 80% of Claude Code’s system prompt for the Claude 5 generation, suggesting the format is tightly coupled to the model.

The “split brain” problem in practice

Lütke’s concern is not theoretical. Shopify runs a large monorepo and a multi-tool AI stack. Some engineers use Codex, others Cursor, others Claude Code. If Codex reads AGENTS.md and Claude Code reads CLAUDE.md, the same code path can be subject to different constraints.

Imagine a payments module with a strict rule: “Run the full test suite before touching billing; never edit database migration files directly.” That rule lives in AGENTS.md. Codex sees it. Claude Code does not. The result is not a crash; it is a silent divergence in behavior that only shows up at code review, or worse, in production.

Developers have already created workarounds. The most common is a symlink: ln -s AGENTS.md CLAUDE.md. Another is to put @AGENTS.md at the top of CLAUDE.md, which Claude Code will then load. Both work for a single repo.

Lütke’s rebuttal is that these are hacks, not solutions. Agent configs and Claude configs recurse down the directory tree. In a monorepo with thousands of developers, every directory must keep both files in sync. A missing file in one subtree means a team suddenly loses its project context. He called it a “stupid complexity tax” that should not fall on developers.

Why Shopify is in a position to push

Shopify is not a casual observer. Last year Tobi told employees that “proficiency with AI is now a baseline expectation.” AI usage is part of performance reviews, and teams must prove a task cannot be automated before requesting more headcount. The company also open-sourced Dispatch, an agentic security scanning system that runs multiple agents against the same code and cross-checks results with different models.

That architecture makes Shopify naturally multi-agent. And in a multi-agent shop, the configuration layer is the real infrastructure. Lütke’s summary: “The framework that orchestrates agents is what survives. Models change; our rules, processes and security boundaries must stay.”

This is why a file name matters. If every model vendor insists on its own config format, every tool swap forces a rewrite of the organization’s institutional memory. Lütke wants one portable rulebook.

Where this is heading

The immediate winner is AGENTS.md. With more than 60,000 open-source projects already using it, a neutral foundation behind it, and the rest of the major tools on board, it is becoming the default interlingua for coding agents. Anthropic’s eventual support is almost certain; the question is only how deep it will go.

The deeper shift is the commoditization of the model and the rising value of the control plane. As coding agents become an industrial pipeline — running for hours, consuming tokens, modifying files, executing tests — the constraints that govern them matter more than the raw completion quality. Enterprises will choose tools that plug into their existing governance, security and workflow layers, not tools that require a separate rulebook.

This is the same pattern that turned Kubernetes into the winning abstraction for containers and HTTP into the winning protocol for the web. A neutral, vendor-agnostic control layer outlasts any single product. For coding agents, that layer is now being defined.

What teams should do now

• Adopt AGENTS.md as the single source of truth. Put project structure, build commands, test commands, lint rules, security constraints and prohibited operations in one file at the repo root.
• Bridge Claude Code today. Add @AGENTS.md to the top of CLAUDE.md so Claude loads the shared rules while keeping Claude-specific tweaks separate.
• Keep vendor-specific files thin. Treat CLAUDE.md, .cursorrules or other formats as thin overlays on top of AGENTS.md, not parallel rulebooks.
• Audit for split brain. In monorepos, scan subdirectories to make sure no agent is silently missing its rules. A missing config is a silent risk.
• Design for model rotation. If your security or quality process depends on a single model reading a specific file, you are already locked in. Build workflows that can swap models without swapping policy.

The real lesson from Shopify is not that Claude Code is bad. It is that the next competitive moat in enterprise AI is not model intelligence but model interoperability — and the humble Markdown file at the root of your repo is becoming the border checkpoint.