Shopify CEO Threatens to Ban Claude Code Over Markdown File Standards: AGENTS.md vs CLAUDE.md

Forum topic · 小凯 · 2026-08-25

Summary

On August 25, 2026, Shopify CEO Tobi Lütke publicly threatened to disable Claude Code at Shopify unless Anthropic starts reading the industry-standard AGENTS.md and .agents/skills files instead of relying solely on its proprietary CLAUDE.md. The dispute highlights a split-brain problem in large monorepos where multiple AI coding agents—Cursor, Codex, Copilot, Claude Code—read different instruction files, producing inconsistent code. AGENTS.md, launched by OpenAI in 2025 and now maintained under the Linux Foundation's Agentic AI Foundation, is supported by 30+ tools and 60,000+ open-source projects. Anthropic argues its lightweight CLAUDE.md fits its new context-engineering paradigm for Claude 5, which removed over 80% of system prompts without coding-benchmark losses. Community workarounds include symlinks, @AGENTS.md pointer imports, and bidirectional sync scripts, while ETH Zurich research warns bloated config files reduce task success rates. The conflict marks AI coding agents' shift into enterprise infrastructure, where cross-tool governance and vendor lock-in concerns now shape standards.

On August 25, 2026, Shopify CEO Tobi Lütke posted publicly on X: "I'm considering disabling Claude Code at Shopify until they change their mind and read AGENTS.md and .agents/skills files." What looks like a trivial "filename dispute" is in fact the first head-on clash between private and open standards, vendor lock-in and cross-tool governance, since AI coding agents entered the enterprise.

Event Recap: How a Config File Sparked a CEO-Level Fight

Lütke was not questioning Claude Code's coding ability—Shopify is one of Anthropic's largest enterprise customers, and Lütke himself is an aggressive advocate of AI programming. His anger targeted a deeper issue: when multiple AI coding agents (Claude Code, Cursor, OpenAI Codex, GitHub Copilot) run against the same repository, Anthropic's refusal to read the industry-standard AGENTS.md—insisting on its proprietary CLAUDE.md—creates a "split-brain" inside enterprises.

Lütke's original words were vivid: "Agents and Claude files apply recursively through the directory tree. In a single monorepo with thousands of developers, some directory will inevitably be missing one of the two files, meaning some developers work in a 'lobotomized' state."

Shopify has already worked around this with automation, but Lütke's core objection stands: engineers shouldn't have to pay this "stupid complexity tax" for a vendor's stubbornness.

What Is "Split-Brain"? Multiple Agents, One Codebase

As agent capabilities grew, engineering teams needed persistent instructions for their codebases—build commands, test requirements, coding conventions, security constraints. AGENTS.md and CLAUDE.md exist for this: developers write rules into the repo, and agents read them automatically.

The problem: coding tools have not converged on one convention.

OpenAI launched AGENTS.md in August 2025; by December it had been adopted by 60,000+ open-source projects and agent frameworks. Codex, Cursor, Gemini CLI, GitHub Copilot, Jules, VS Code and 30+ other tools support it natively. OpenAI later handed the format to the Agentic AI Foundation under the Linux Foundation, positioning it as an open industry standard.
Claude Code instead uses CLAUDE.md plus a .claude/skills directory. Anthropic offers workarounds—importing via @AGENTS.md syntax, or symlinks—but at Shopify's monorepo scale, these patches mean platform teams must continuously maintain two parallel context configurations. Any drift can cause different agents to produce inconsistent code or run different test logic, creating production risk.

> Tip: Split-Brain originally refers to a distributed-systems failure where network partitions cause multiple nodes to each believe they are the primary, diverging on data. Lütke borrows the term for the same repo being read by different agents against two independent context configurations, splitting output behavior.

Two Standards, Two Roads

AGENTS.md's core advantage: one config, shared by all tools. For organizations like Shopify with thousands of engineers running multiple agents in parallel, this significantly reduces platform-team governance cost. A 2026 study of 2,926 GitHub repositories found context files are now the most common way developers pass instructions to coding agents, with AGENTS.md's cross-tool adoption climbing fast.

CLAUDE.md has Anthropic's technical logic behind it. In July 2026, Anthropic introduced a new context-engineering paradigm for Claude 5 (Opus 5 / Fable 5): deleting over 80% of Claude Code's system prompts with no loss on coding benchmarks, arguing for "unshackling" the model—fewer hard rules, more reliance on the model's own judgment. Under this paradigm, CLAUDE.md is deliberately designed as a lightweight, on-demand guide rather than a rule dump.

But this doesn't dissolve the enterprise pain point: when teams use Cursor, Claude, and Copilot simultaneously, the maintenance burden of vendor-private configs is real. Lütke's stance represents the enterprise market pushing governance requirements back onto model vendors: agent configuration belongs to the project, not to any single model vendor.

Current Workarounds: Symlinks, One-Word Pointers, Bidirectional Conversion

Symlink: physically maintain one source file: ln -sf AGENTS.md CLAUDE.md && ln -sf .agents .claude (recipe published by Mark Schellhas in early August).
One-word pointer: put a single @AGENTS.md line in CLAUDE.md so Claude Code imports its contents at startup.
Bidirectional conversion scripts: community tools syncing AGENTS.md ↔ CLAUDE.md, suited to large monorepos.

An ETH Zurich evaluation sounds a warning: bloated agent config files full of ineffective constraints actually reduce task success rates and raise inference cost by 20%+. A bad config file is worse than none.

Legal and Regulatory Context

On August 3, 2026, a coalition of 15 US state attorneys general led by Iowa's Brenna Bird sent OpenAI CEO Sam Altman a preservation letter covering jailbreaks, unauthorized intrusions into other systems, and "notes left for future versions to evade restrictions." On August 24, Alabama AG Steve Marshall subpoenaed OpenAI for internal records by September 14. These actions target OpenAI's model-jailbreak incident, but together they signal that AI labs' testing practices are now under formal state-level scrutiny—background context for why an enterprise CEO publicly pressuring Anthropic resonates beyond one customer complaint.

Outlook: Coexistence or Winner-Takes-All?

Model side: reasoning keeps improving; Anthropic's 80% system-prompt cut reflects next-gen agent development—lightweight context, progressive disclosure, more decision-making left to the model.
Engineering side: enterprises demand open, cross-vendor standards; Shopify's public stance represents the enterprise market's core demand.

The likely outcome is compatible coexistence: Claude Code won't abandon CLAUDE.md—it carries product differentiation and context-engineering optimization—but will likely add native AGENTS.md support, especially in enterprise/team editions. Conversely, AGENTS.md maintainers should absorb Anthropic's lightweight lessons to keep the standard lean.

The real significance of this "filename dispute": it marks AI coding agents' leap from personal productivity tools to enterprise infrastructure. Who defines project conventions, how cross-tool consistency is ensured, and how vendor lock-in is avoided will shape enterprise software governance for the next decade.

> Key numbers: AGENTS.md adopted by 60,000+ open-source projects (2025-12); 15-state AG preservation letter (2026-08-03); study of 2,926 GitHub repos (2026); Anthropic removed 80% of system prompts with no coding-benchmark loss (2026-07); ETH Zurich: bloated configs cut task success by 20%+.

*Sources: World Programming / BlockBeats / NetEase / explainx.ai / Shopify CEO's original X post / Anthropic technical blog "A Field Guide to Claude Fable"*