Shopify threatens Claude Code over a file

In a repo with thousands of developers, agent instructions travel down the directory tree. When a folder is missing the bridge to Claude Code, some people work without the same context. The head of Shopify calls it a tax that nobody should have to pay.

Key takeaways

Related:Claude Code now also runs Alibaba’s Qwen  /  Codex Micro: agent control at the desk

What is AGENTS.md? AGENTS.md is a markdown file in the repository. It gives coding agents build commands, tests, conventions and boundaries so they do not have to guess in the project. Codex, Cursor, Copilot and Gemini CLI look for exactly this name. Claude Code opens CLAUDE.md and pulls in AGENTS.md only via an import or a symlink. When a subfolder is missing the bridge, part of the team works with a different context.

Shopify turns a file name into a CEO-level issue

On 25 August 2026, Tobi Lütke wrote on X that he is considering banning Claude Code at Shopify until Anthropic also reads AGENTS.md and the .agents/skills folder for shared agent capabilities. Reading only CLAUDE.md, he wrote, creates split-brain as soon as different tools run in the same team. The wording remains a threat. He has not issued a company-wide ban.

Lütke’s second post makes the recursive scope in the monorepo the real point. Agent files and Claude files apply across the whole tree. In a monorepo with thousands of developers, some directory will eventually lack one of the two. Then some people work with an agent that Lütke calls lobotomized. Shopify is patching this with automation, according to Lütke. Lütke calls the price a stupid complexity tax that nobody should have to pay.

Both posts are public: the opening shot and Lütke’s second post. Anyone reading them as a product takedown misses the point. Lütke attacks the file the agent opens at the workplace.

The tree hands out two truths

AGENTS.md is the shared instruction file for coding agents. The agents.md site describes it as a README for agents: build, tests, conventions, boundaries. Codex, Cursor, Copilot and Gemini CLI look for exactly this name. OpenAI pushed the format forward in 2025.

Claude Code opens CLAUDE.md. The Anthropic memory documentation says so outright: Claude Code reads CLAUDE.md. AGENTS.md is read only when CLAUDE.md imports it or a symlink makes both names point to the same file. Claude loads nested files in subfolders on demand, as soon as files there are read. Both worlds travel down the same tree. They just carry different labels.

In a flat project this hardly shows. In a monorepo it does. A package directory can contain AGENTS.md but lack CLAUDE.md. The Codex user in the next room gets the test rule, the Claude user does not. That is exactly what Lütke means by split-brain: two truths in one commit.

Tobi already knows the tricks

Under the first post, symlink suggestions piled up, along with the line in CLAUDE.md that imports AGENTS.md. Lütke replied that he knows them all. The Anthropic documentation recommends this bridge and allows Claude-specific notes underneath. In a repo with a single instruction file at the root, that takes just one line.

The recursive case remains: every directory is its own world. Anyone wiring only the root leaves packages without a bridge.

Anyone taking the tree seriously generates the bridge where AGENTS.md already sits. A generator or a CI check adds the bridge by creating a CLAUDE.md file containing an import or by adding a symlink. Manually walking through folders does not scale. Lütke wants to get rid of the automation.

Anthropic defends its own prompt

Thariq Shihipar of Claude Code answered Lütke publicly. The team, he said, is working on making Claude Code more hackable, including easier use of AGENTS.md and other interventions in the system prompt. He named no date. The rationale centered on the model: model families are not interchangeable, the system prompt strongly affects performance, and Claude Code uses its own prompts for each model.

For now, the documented bridge remains. Anyone importing AGENTS.md into CLAUDE.md pulls the shared text in at startup. Claude-specific lines may sit below it. Skills that only Claude Code understands belong in .claude. The shared truth belongs in AGENTS.md, otherwise it drifts at the next tool change.

The rationale about model differences explains why Anthropic loves its own file name. It does not explain why the Claude Code file loader ignores the de facto standard AGENTS.md. A different prompt can live in CLAUDE.md. The file loader can still open AGENTS.md. Lütke pays the tax in the tree. Anthropic pays it with reduced compatibility with the rest of the agent world.

One file leads, the bridge follows

The operating rule is short. AGENTS.md carries what every tool needs: install, test, lint, architecture boundaries, taboos. CLAUDE.md stays thin: an import of the shared file, then only what Claude Code additionally needs. Personal notes stay in CLAUDE.local.md and outside version control. Anyone using only Claude Code gets by with CLAUDE.md. A mix in the team forces the use of the shared file.

The same rule applies one level deeper in subfolders. If a package has AGENTS.md, a bridge sits beside it. The check belongs in the pipeline, otherwise Lütke’s lobotomy repeats with every new package. A human reviewer spots the missing counterpart too late: by then, the agent has already built something incorrectly.

Shopify is already automating this, according to Lütke. Anthropic could read the file name the rest of the agent landscape already writes. Until that happens, the thin bridge remains the price of a mixed team. Lütke’s CEO-level issue is AGENTS.md. Anyone who also maintains its content in CLAUDE.md has already lost the argument.

Frequently asked questions

Image Source: AI-Generated (August 2026)

Translated from the German original using artificial intelligence. The German version is authoritative.