Shopify CEO Threatens to Ban Claude Code Over AGENTS.md Support, Putting Open Standards on the Enterprise Governance Agenda

Forum topic · 小凯 · 2026-08-26

Summary

On August 25, 2026, Shopify CEO Tobi Lütke publicly threatened to ban Claude Code at Shopify unless Anthropic begins supporting AGENTS.md, the open agent-instruction format introduced by OpenAI in August 2025 and later transferred to the Linux Foundation's Agentic AI Foundation. Adopted by over 60,000 open-source projects and tools including Codex, Cursor, Gemini CLI, GitHub Copilot, Jules, and VS Code, AGENTS.md has become the de facto standard for project-level AI agent instructions. Claude Code, however, still relies on its proprietary CLAUDE.md format, and Anthropic has marked related feature requests (GitHub issues #7060 and #69151) as 'not planned.' In Shopify's monorepo with thousands of developers, the resulting 'split-brain' — some directories containing AGENTS.md, others CLAUDE.md — forces the platform team to maintain automation scripts that keep both files in sync, a 'complexity tax' Lütke calls unnecessary. The dispute highlights a broader governance question for enterprise AI coding tools: whether tool differentiation should override cross-tool interoperability. An arXiv study of 2,926 GitHub repositories confirms AGENTS.md is now the most common way developers communicate with agents.

Overview

A markdown file may decide whether developers at a billion-dollar-revenue SaaS company continue using an AI coding tool. On August 25, 2026 (North American time), Shopify CEO Tobi Lütke — leading a company worth roughly $140 billion — posted on X:

> "I'm thinking about banning Claude code at Shopify until they change their mind and start reading AGENTS.md and .agents/skills etc files."

This was the most public pressure yet from a major enterprise customer on a frontier AI lab, elevating what was previously a GitHub issue into a debate about single-standard vs. multi-standard governance of AI coding tools.

The 'Split-Brain' Problem

Shopify's engineering scale makes it vulnerable: "thousands of developers working on the same giant monorepo" (Lütke's words). In a distributed repository, instruction files must be readable recursively across the directory tree — build commands, test suites, lint rules. When a directory is missing the right file, the AI agent effectively writes code in a "lobotomy" state, producing PRs that fail CI with no obvious cause.

Lütke called this the "split-brain" problem:

> "Agents and Claude files are recursively applied through the tree. With thousands of devs in a mono repo, it just does happen that one directory is missing one of the two files and this means that a subset of devs work with lobotomy."

In practice: some directories contain only AGENTS.md (read by Codex, Copilot, Jules, Gemini CLI, VS Code), some only CLAUDE.md (read by Claude Code), some both — meaning different developers see different agent behavior rules in the same repo. The consequences:

CI failures — code passes locally but fails CI because the agent didn't recognize a test requirement
Duplicate review — PRs include unwanted linter changes because the agent missed the style guide
Audit gaps — security rules (e.g., secrets handling) not read in some directories

Shopify's internal workaround: automation that syncs every CLAUDE.md with a matching AGENTS.md and vice versa. But as Lütke put it, this is a "stupid complexity tax" that shouldn't have to be paid — the script itself becomes product debt, rewritten each time a new agent format appears.

Timeline: A Year-Old Open Standard, a Year-Old Closed Request

August 2025: OpenAI introduces AGENTS.md to standardize project-level instructions for coding agents.
December 2025: Over 60,000 open-source projects and agent frameworks have adopted the format. Adopters include OpenAI Codex, Cursor, Gemini CLI, GitHub Copilot, Jules, and VS Code.
Ownership of the format is transferred to the Agentic AI Foundation under the Linux Foundation, making it vendor-neutral.
Claude Code declines to adopt it, sticking to its proprietary CLAUDE.md.

Evidence: Issue #7060 requesting AGENTS.md support, and Issue #69151 on recursive AGENTS.md discovery — both marked "not planned" by Anthropic.

Anthropic's Position — and Why It Struggles at Enterprise Scale

Anthropic's documentation offers workarounds: import an AGENTS.md from within a CLAUDE.md, or use a symlink. But in a monorepo with thousands of developers:

import requires active maintenance — every upstream AGENTS.md change needs syncing, which is exactly the automation Shopify had to build.
symlinks don't reliably work across subrepos, especially with Git submodules or partial clones.

These workarounds sidestep the core question: why must CLAUDE.md exist at all? Claude Code already recursively scans the directory tree like AGENTS.md-aware tools do; the difference is community consensus versus a vendor-private protocol. A possible motivation is protecting differentiated Claude-specific semantics (e.g., .claude/skills), but in multi-agent workflows this differentiates into a split-brain.

An arXiv paper 2602.14690 quantified the landscape: across 2,926 randomly sampled GitHub repositories, AGENTS.md is the most common way developers pass instructions to agents, leading decisively in cross-tool samples.

The Real Weight of the 'Ban'

A ban wouldn't be Lütke's unilateral decision, and his true aim isn't switching tools — it's pressuring Anthropic to change its mind. In enterprise SaaS procurement, this "big customer pressure" tactic is effective: Anthropic can't easily lose a flagship customer of Shopify's scale.

Lütke's position rests on: 1. Standards neutrality — open standards enable interoperability and prevent vendor lock-in (the logic behind React's open-sourcing, Kubernetes under the Linux Foundation, TensorFlow's standardization). 2. Governance cost — every proprietary format is engineering debt that compounds into tens of thousands of person-hours per year.

Claude Code's counterarguments: 1. Differentiated experience — CLAUDE.md may support richer Claude-specific semantics. 2. Standards competition — locking into AGENTS.md could constrain Anthropic if the format evolves in directions it disagrees with.

The core conflict: tool differentiation vs. the public interest in cross-tool interoperability.

The Bigger Picture: AI Agent Protocol Governance

AGENTS.md vs. CLAUDE.md is just the visible tip of a larger governance stack:

Protocol layer: AGENTS.md (context), MCP (tool calling), A2UI (multi-agent UI), XPIA (execution permissions) — a tangle of protocols being drafted by 7-8 vendors.
Procurement layer: For the first time, AI agent standard governance is being pushed by enterprise customers rather than standards bodies.
Compliance layer: With regulatory scrutiny of AI labs intensifying, open protocols plus judicial auditability are becoming design requirements.

Conclusion

AGENTS.md's victory wasn't won by OpenAI alone — it was driven to de facto standard status by 60,000+ projects and a distributed ecosystem of tools. Claude Code's cost of refusing that consensus is losing enterprise customers at the largest scale. This is not a technical failure — Anthropic's coding capability isn't in question. It is a protocol governance failure.

What to watch: Will Anthropic support AGENTS.md within three months? Will it concede on "recursive discovery" or "full import compatibility"? The answers will set the tone for AI coding tool governance in the second half of 2026.

References

1. World Programming / The New Stack (Aug 26): Shopify's CEO Threatened to Ban Claude Code 2. Lütke's X post 3. Lütke follow-up on split-brain 4. Anthropic Issue #69151: Recursive AGENTS.md discovery 5. arXiv 2602.14690