Why Shopify’s CEO Is Considering Disabling Claude Code Over AGENTS.md Incompatibility

Shopify CEO Tobi Lütke is contemplating disabling Claude Code until it supports reading AGENTS.md and .agents/skills files. He argues that Claude Code’s insistence on reading CLAUDE.md can give developers using different AI coding tools divergent project rules and instructions within the same repository.

As AI coding agents become part of software development workflows, many repositories now include dedicated AI‑specific description files that capture project structure, coding standards, test requirements, development processes, and callable skills. Files such as AGENTS.md have become a common convention, with tools like OpenAI Codex, Cursor, and Amp already supporting the format. Claude Code, however, uses a separate set of files— CLAUDE.md and .claude/skills —to store similar information.

When team members use multiple AI coding tools, the two configuration entry points can diverge. For example, a developer using Codex reads AGENTS.md, while a developer using Claude Code reads CLAUDE.md. If the files differ in any directory, the AI agents receive different project rules.

The developer community quickly offered practical solutions. The most common method is to create a symbolic link in the project: ln -s AGENTS.md CLAUDE.md Another approach leverages Claude Code’s file‑reference capability. By inserting @AGENTS.md inside CLAUDE.md, Claude Code can load the shared rules while still retaining space for Claude‑specific configuration.

Lütke later responded to the flood of suggestions, noting that both agent and Claude configuration files are applied recursively along the directory tree. In a large monorepo with thousands of developers, it is difficult to ensure that every directory maintains both files in sync. Missing one file can cause developers using different AI tools to see different project contexts, which he describes as a “complexity tax” that developers should not have to bear.

Anthropic’s Claude Code team acknowledges the issue. Team member Thariq explained that different model families have distinct behavior, and system prompts heavily influence performance. Claude models prefer particular organization of skills, system prompt, and CLAUDE.md. The team is working to make Claude Code more flexible, aligning with recent discussions on “context engineering.” They plan to continue sharing progress, including short‑term solutions that reference other markdown files (e.g., @AGENTS.md) to load existing content, and they admitted that supporting multiple configuration sets adds maintenance overhead.

Overall, while the proposed fixes are straightforward for small projects, Lütke’s concern centers on the long‑term maintenance challenges in a massive monorepo like Shopify’s, where configuration files must evolve alongside directory structures, project boundaries, and team conventions.