AGENTS.md Firestorm: Shopify CEO Threatens to Ban Claude Code
> Originally posted in Chinese on zhichai.net, August 26, 2026 · AI Coding Vendor Governance
The story in one sentence
A single markdown file now determines whether developers at a billion-dollar-revenue SaaS company can keep using an AI coding tool.
On August 25, 2026, Shopify CEO Tobi Lütke (company valued around $140B) 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." It is the most blunt public pressure yet from a major e-commerce SaaS vendor against a frontier model company — elevating a closed GitHub feature request into a public governance fight.
Background: AGENTS.md is an open protocol introduced by OpenAI in August 2025 so coding agents (Codex and others) can read project-level instructions. By December 2025, over 60,000 open-source projects and agent frameworks had adopted it, and OpenAI transferred ownership of the format to the Agentic AI Foundation under the Linux Foundation. Claude Code, however, still insists on its proprietary CLAUDE.md format — despite nearly a year of developer requests.
> "We fix this with automation," Lütke said, "but it's a complexity tax that shouldn't be paid."
Why the harm is real: the split-brain incident
Shopify's engineering scale makes it fragile: "thousands of devs in a mono repo" (Lütke's words). Instruction files must be discovered recursively across the directory tree — build commands, test suites, lint rules. If a file is missing or misplaced, the agent writes code in a "lobotomy" state; PRs 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 plain terms: 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 — so different developers on the same team see different agent behavior rules.
Note: in distributed systems, "split-brain" describes cluster nodes acting inconsistently while each believes it is correct. Lütke borrows the term for developers whose AI tools each follow different "rules" in the same codebase.
Concrete consequences:
- Red CI: local passes, CI fails because the agent missed a test requirement.
- Duplicate reviews: PRs include unneeded linter changes because the agent never read the style guide.
- Audit gaps: a security agent misses key-handling rules in one directory and commits secrets in plaintext.
- Aug 2025: OpenAI introduces AGENTS.md to turn "project instructions for agents" into a shared standard instead of per-tool filenames. Motivation: developers were copying the same instructions 5–7 times for 5–7 agent tools.
- Dec 2025: adoption exceeds 60,000 projects; compatible tools include Codex, Cursor, Gemini CLI, GitHub Copilot, Jules, VS Code. OpenAI transfers the format to the Agentic AI Foundation (Linux Foundation) — moving ownership from one company to a neutral foundation.
- Claude Code refuses to read it. Evidence: Issue #7060 requesting AGENTS.md support, and the more recent Issue #69151 on recursive AGENTS.md discovery. Both were closed by Anthropic as "not planned".
importneeds active maintenance: every upstream change must be mirrored — exactly what Lütke's "automation" watcher does.- symlinks break across subrepos: Git submodules and partial clones make cross-module symlinks unreliable.
- Protocol layer: AGENTS.md (context), MCP (tool calling), A2UI (multi-agent UI), XPIA (execution permissions) — an interweaving protocol stack being drafted separately by 7–8 vendors.
- Procurement layer: for the first time, agent standard-setting is being pushed by enterprise customers rather than standards bodies.
- Compliance layer: a multi-state attorneys-general preservation letter to major AI labs taking effect this week suggests agent protocols will soon need "compliance by design" — open protocols plus judicial auditability.
Shopify's internal fix: a platform-team script that generates a matching AGENTS.md for every CLAUDE.md and vice versa. Shopify hasn't actually fallen over, but the script itself is product debt — a patch for a tooling defect that must be rewritten each time a new agent format appears.
Lütke's real anger: not that Claude Code is bad, but that Anthropic refuses to acknowledge an open standard that has existed for a year across multiple tools.
Timeline: a one-year-old open standard, a one-year-old closed request
Anthropic's workarounds — and why they don't scale
Anthropic's docs offer two workarounds: import AGENTS.md from CLAUDE.md, or symlink the two files. In a monorepo with thousands of developers:
Both answers dodge the core question: why must CLAUDE.md exist at all? Claude Code already recursively scans the tree, much like AGENTS.md. The difference is at the community-consensus level, not the engineering level: one format is used by 60,000+ projects, the other by one vendor's tool alone. Anthropic's likely rationale — protecting differentiated Claude-specific semantics (e.g., .claude/skills) — makes sense for single-agent use but collapses in multi-agent workflows, where differentiation degrades into split-brain.
Independent empirical support: an August 24 arXiv paper (2602.14690) analyzed 2,926 randomly sampled GitHub repositories and found AGENTS.md is already the most common way developers pass instructions to agents across tools.
The real weight of the "ban"
Lütke can't ban Claude Code unilaterally; many of Shopify's thousands of engineers use it daily. The threat is really a procurement-leverage move: "I'm not hitting you with my left hand; I'm using 'I might leave' to force you back to the table." Anthropic cannot afford to lose a flagship customer at Shopify's scale. Notably, Anthropic launched its Anthropic Economic Index the same day — possibly unrelated, but its密集 public outreach that week signals it is catching up on enterprise tool governance.
Both sides have defensible positions:
Lütke / open standards: 1. Neutral standards enable interoperability and prevent vendor lock-in — the same logic as React's open-sourcing, Kubernetes under the Linux Foundation, TensorFlow's standardization. 2. Every proprietary format is engineering debt; summed across tools it becomes tens of thousands of hours per year.
Claude Code: 1. CLAUDE.md may allow richer semantics (private skill calls, context-window tuning). 2. If AGENTS.md evolves in directions Anthropic disagrees with (e.g., Microsoft-backed XPIA or Google-backed A2UI proposals), it would be locked into someone else's protocol.
The core tension: tool differentiation vs. the public good of cross-tool interoperability.
The bigger open question: AI agent protocol governance
Bottom line
AGENTS.md's win wasn't OpenAI's alone — it was pushed to de facto standard status by 60,000 projects plus Cursor, Gemini CLI, Copilot, Jules, and VS Code. Claude Code's cost: refusing a broad consensus risks a frontier model company losing its largest-tier enterprise customers. This isn't a capability failure — Anthropic's coding ability is not weak. It is a protocol governance failure.
What to watch: will Anthropic endorse AGENTS.md within three months? Will it concede recursive discovery, full import compatibility, or more? 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 — Anthropic had already closed the feature request 2. Lütke's original X post: https://x.com/tobi/status/2092259436538495186 3. Lütke's follow-up on split-brain: https://x.com/tobi/status/2092308116524138711 4. Anthropic Claude Code Issue #69151 (recursive AGENTS.md discovery): https://github.com/anthropics/claude-code/issues/69151 5. arXiv 2602.14690 — statistics on context-file usage in GitHub repositories: https://arxiv.org/abs/2602.14690