Shopify CEO Tobi Lütke said publicly on August 25 that he's considering banning Claude Code inside the company unless Anthropic changes course and has it read files like AGENTS.md and .agents/skills. His reasoning: Claude Code's insistence on reading only CLAUDE.md causes a "split brain" when team members use different tools, a complexity tax he says nobody should have to pay.
The Fight Is Over One Markdown File
AGENTS.md is a convention OpenAI introduced in August 2025: build commands, test instructions, and security rules for a project go into a single Markdown file at the repo root, which coding agents read to understand the codebase. By the end of 2025, public data showed more than 60,000 open-source projects had adopted the file, and the standard itself was later handed off to the Agentic AI Foundation, a project under the Linux Foundation. Codex, Cursor, Gemini CLI, GitHub Copilot, Jules, and VS Code all recognize it.
Claude Code, meanwhile, uses its own CLAUDE.md. For a solo project, that's just one extra file to write. For an organization like Shopify, where thousands of engineers share large repositories, the problem changes shape: if a directory only has one of the two files, a colleague using a different tool ends up working with incomplete context.
A Request Left Open for a Year
The GitHub issue asking Claude Code to natively support AGENTS.md has been open since August 2025 and has drawn nearly 5,000 upvotes, making it one of the most-supported compatibility requests in the repo. Several similar issues were closed one after another, one with the verdict "no plans at this time." Developers left hundreds of comments underneath, but there was never an official statement.
The community had long since worked out its own fixes: turning CLAUDE.md into a symlink pointing at AGENTS.md, or simply writing a line in CLAUDE.md telling the tool to go read AGENTS.md instead. It works, but it has to be redone in every repository.
Anthropic Finally Gives a Reason
On the same day Lütke spoke up, Anthropic's Thariq replied, saying the team is making Claude Code more customizable so it can easily use AGENTS.md or let users edit the system prompt directly, with more details to come once it's ready. He also gave, for the first time, a reason for the earlier refusal: Anthropic doesn't believe different model families are interchangeable, and system prompts have a major effect on performance — Claude Code maintains separate system prompts per model internally, which is already a significant maintenance burden.
The reasoning holds up technically. Feeding the same instructions to models from different vendors really does produce noticeably different results, and prompts tuned for a specific model can open up a clear performance gap. But that explains why CLAUDE.md should exist on its own — it doesn't explain why Claude Code can't also read AGENTS.md while it's at it. The two aren't in conflict.
Context Files Become the New Battleground
A survey covering 2,926 GitHub repositories found that context files are already the most common way developers give instructions to coding agents. Once that slot is claimed by one vendor's proprietary format, the migration cost falls on users — today it's one extra file to maintain, tomorrow it could be skills, sub-agents, and permission configs all locked into one company's directory structure.
Measured in engineering effort, the standoff looks disproportionate. Having Claude Code fall back to AGENTS.md when it can't find CLAUDE.md is a few dozen lines of file-lookup logic — an entirely different order of magnitude from maintaining separate system prompts per model. If the concern is that generic instructions might drag down performance, a single line of caveat text would cover it. That's exactly the compromise the community kept proposing over the past year: read CLAUDE.md by default, fall back to AGENTS.md if it's missing, and let CLAUDE.md take precedence when both exist.
Why Lütke's public pressure worked is simple enough. Shopify is a large-scale user of Claude Code, and a public-company CEO threatening to ban a tool carries far more weight than 5,000 anonymous upvotes. For Anthropic, Claude Code's current edge comes from product experience, not from locking users into a file format — and at a moment when developers routinely juggle multiple tools, turning compatibility into a simple toggle is a better trade than defending one file name.
Sources: The New Stack, GitHub, CocoLoop, Hacker News. Adoption figures for AGENTS.md, the issue's upvote count, and Anthropic's statement were verified against publicly available pages.