Why Shopify’s CEO Is Threatening to Ban Claude Code Over AGENTS.md

Shopify’s threatened Claude Code ban is a standards and operations dispute disguised as a configuration-file argument. Tobi Lütke says Claude Code’s insistence on its own CLAUDE.md convention can create “split brain” behavior when engineers use several AI coding tools against the same repository. His position is that a large engineering organization should maintain one shared set of agent instructions, not duplicate it for each vendor. 415

What Shopify is objecting to

AGENTS.md is a plain Markdown file that gives coding agents project-specific context: how to build and test the code, which conventions to follow, and what boundaries or contribution rules apply. The format is designed to live with the repository, so the instructions travel with the code rather than with a particular developer’s tool. 2931

Claude Code’s documented project context instead centers on CLAUDE.md, .claude/rules, project skills, hooks, and settings. Its documentation describes files discovered in the current directory and parent directories, creating a recursive instruction system of its own. 2

That distinction is manageable for a solo developer. It becomes more consequential in a large monorepo where different teams may use Claude Code alongside Codex, Cursor, Gemini CLI, or GitHub Copilot. If one tool reads AGENTS.md and another reads only CLAUDE.md, the same directory can present different build commands, coding rules, or “do not modify” boundaries to different agents.

Why the monorepo problem is bigger than duplicate documentation

The practical choices for a company using multiple agents are unattractive:

• maintain both AGENTS.md and CLAUDE.md at the repository root;
• repeat that duplication in relevant subdirectories;
• keep the two files synchronized as the codebase changes; and
• add checks or tooling to detect when the instructions drift apart.

A missed update is not merely a documentation defect. One agent may follow a new test command while another uses an obsolete one. One may see a restriction on a directory while another does not. The resulting behavior can produce inconsistent edits, reviews, and developer expectations—the “split brain” problem Lütke highlighted. 47

The public Claude Code feature request describes the immediate friction directly: when a repository contains AGENTS.md, Claude Code may ignore that context, leaving the developer to create a CLAUDE.md file and duplicate the instructions. 12

How AGENTS.md became an interoperability flashpoint

AGENTS.md began as an open convention associated with OpenAI’s Codex ecosystem and was developed amid broader industry collaboration involving companies such as Google, OpenAI, Sourcegraph, Factory, and Cursor. 1931

On December 9, 2025, OpenAI and Anthropic contributed AGENTS.md and other agent infrastructure to the Linux Foundation-backed Agentic AI Foundation. The goal of the foundation is neutral stewardship for open, interoperable agent technology rather than control by one commercial tool vendor. 2829

The format’s adoption gives the dispute practical weight. The Linux Foundation says AGENTS.md had already been adopted by more than 60,000 open-source projects and agent frameworks, including Codex, Cursor, Devin, Factory, Gemini CLI, GitHub Copilot, Jules, and VS Code. 31

That does not make AGENTS.md a mandatory technical standard. It does make it a widely used portability layer: teams can place durable repository guidance in one visible, version-controlled location and expect multiple tools to find it.

What Claude Code offers instead

The documented workaround is to make Claude Code load shared instructions through its native configuration. One approach is to reference AGENTS.md from CLAUDE.md; another reported approach is to use a symbolic link or other configuration bridge. 69

Those methods can reduce the immediate incompatibility, but they do not provide the same experience as native support. They add a wrapper, require repository-specific setup, and leave teams responsible for deciding how imports, precedence, directory scope, and updates should work. Symbolic links can also interact differently with operating systems, Git settings, IDEs, containers, and enterprise security policies.

Anthropic’s public response, as quoted in the discussion, says the company is working to make Claude Code more configurable, including making it easier to use AGENTS.md or modify other system-prompt settings. 15 The available material does not establish a definitive technical or commercial reason why Claude Code has not simply added native recognition, so claims about Anthropic’s motive should be treated as speculation.

The apparent standards contradiction

The dispute is especially notable because Anthropic is associated with the wider Agentic AI Foundation ecosystem, while Claude Code’s documented project conventions remain centered on Anthropic-specific files. 22829

That does not necessarily mean Anthropic opposes open standards. A vendor can support an open baseline while retaining proprietary extensions for capabilities that are specific to its models or product. But from an enterprise buyer’s perspective, the distinction matters less if the core project instructions still need to be rewritten or bridged before another tool can use them.

What the standoff could mean for AI developer tools

A recurring configuration tax

Every vendor-specific instruction tree creates ongoing maintenance work: duplicated files, synchronization checks, migration scripts, and testing. As organizations add more coding agents, the cost grows with the number of tools and repository scopes they support.

More pressure against vendor lock-in

Repository instructions often encode team practices, security boundaries, workflows, and operational knowledge. If those instructions work best in one vendor’s format, switching tools or running several tools in parallel becomes more expensive. A shared format lowers that switching cost and gives engineering leaders more leverage in procurement decisions.

A likely two-layer model

The most practical outcome may not be one format replacing every proprietary convention. A common baseline such as AGENTS.md could hold portable project guidance, while CLAUDE.md or other vendor-specific files provide optional extensions. That arrangement preserves product differentiation without making basic repository context non-portable.

Enterprise buyers become standards setters

Lütke’s response shows how large customers can turn a small compatibility request into a product requirement. If companies such as Shopify make cross-tool instruction support a condition of deployment, vendors may have to treat interoperability as part of enterprise readiness rather than as a community nice-to-have.

Bottom line

Shopify is not threatening to ban Claude Code because CLAUDE.md is inherently inferior to AGENTS.md. The complaint is that a proprietary default becomes costly when many engineers use several agents on the same codebase.

AGENTS.md’s broad adoption, Linux Foundation governance, and support across major coding tools make it a credible shared layer. 2931 Claude Code’s imports and symlinks can serve as workarounds, but they move the integration burden to each organization. The larger question is whether AI coding agents will become interoperable infrastructure—or closed platforms whose project knowledge must be adapted for every tool.