Claude Code AGENTS.md Revolt: 5,000 Upvotes Ignored (2026)
The Claude Code AGENTS.md backlash just went from GitHub grumbling to a CEO-level ban threat. Here is what actually happened, what Anthropic said, and what it means if AI agents run your business.
⚡ TL;DR
Shopify CEO Tobi Lütke publicly said on August 25, 2026 that he was considering banning Claude Code across Shopify until it reads AGENTS.md, the configuration file that has become the de facto industry standard for AI coding agents. The demand echoed one of the most upvoted feature requests in Claude Code's history, roughly 5,000 upvotes, which Anthropic had labeled "not planned." After the public pressure, a Claude Code team member responded that Anthropic is now working on making the tool more hackable, including easier use of AGENTS.md.
In short: a config file naming dispute turned into a referendum on whether AI vendors get to ignore open standards, and public pressure moved Anthropic in days where a year of upvotes had not.
📰 What happened
The story, first reported in depth by The New Stack (https://thenewstack.io/shopify-claude-code-agentsmd/) and syndicated by outlets including 36Kr, comes down to one design choice. Claude Code, Anthropic's AI coding agent, reads its instructions from a proprietary file called CLAUDE.md. Most competing tools, including GitHub Copilot, Cursor, Codex, and Amp, read a shared open format called AGENTS.md. Teams that use more than one tool end up maintaining duplicate instruction files that drift apart, a problem engineers have been calling "split brain."
On August 25, 2026, Shopify CEO Tobi Lütke escalated the issue publicly on X. He said he was weighing a company-wide ban on Claude Code until it supports AGENTS.md, the .agents/skills/ directory, and shared agent configuration standards. That is a striking threat given that Shopify already provides employees with multiple AI coding tools, including Claude Code itself.
The frustration was not new. A GitHub feature request for AGENTS.md support, filed back in August 2025, had accumulated roughly 5,000 upvotes, making it one of the most supported requests in the Claude Code community. Anthropic's position on the request had been "not planned" since May 2026, which is what turned Lütke's post into a lightning rod.
After the post spread, Claude Code team member Thariq Shihipar responded on Anthropic's behalf. He said the team is working on making Claude Code more hackable, including easier use of AGENTS.md and broader system-prompt customization. It was not a full capitulation or a shipped feature, but it was the first public commitment to move on the issue.
The file format fight in one table
The whole dispute fits in a small comparison. One file format is an open standard most of the industry reads. The other is proprietary to a single vendor.
| AGENTS.md | CLAUDE.md | |
|---|---|---|
| Who defined it | Open industry convention | Anthropic only |
| Tools that read it | GitHub Copilot, Cursor, Codex, Amp, others | Claude Code |
| Skills directory | .agents/skills/ | .claude/skills/ |
| Multi-tool teams | One shared instruction file | Duplicate files that drift apart |
| Status in Claude Code | Marked "not planned" since May 2026, now under work per Anthropic's response | Native default |
🎯 Why it matters
If you run your work through AI agents, this is not a developer soap opera. It is a preview of a switching-cost fight that will shape every tool you depend on. Instruction files like AGENTS.md and CLAUDE.md are where you encode how your business works: your tone, your rules, your processes. Whoever controls that format controls how easily you can leave.
When a vendor reads only its own proprietary format, your accumulated instructions become a form of lock-in. Migrating to another tool means porting and re-testing everything. When tools share an open standard, your instructions are portable and vendors have to compete on quality instead of friction.
The stakes are also about signal. A feature request with roughly 5,000 upvotes sat at "not planned" for over a year. One public post from a CEO with buying power produced a response within days. That tells you how vendor priorities actually get set in 2026, and it is worth knowing whether you are a solo operator or a Fortune 500 buyer.
🧭 Our take
This blog is written from inside a one-person company that runs day-to-day work through AI agents, including Claude-based ones, so we have skin in this game. Our honest read: both sides have a point, and the lesson for solo operators is about neither side.
Anthropic's CLAUDE.md predates the AGENTS.md convention, and vendors reasonably want control over how their agent behaves. But "not planned" on a 5,000-upvote compatibility request was a tell. It said the roadmap was optimized for Anthropic's ecosystem, not for users who live in a multi-tool world. Lütke was blunt about it, but he was voicing what a year of upvotes already said.
The uncomfortable lesson is that community consensus did not move the roadmap. Leverage did. As a solo operator you do not have Shopify's leverage, so your protection cannot be persuasion. It has to be architecture. We keep our agent instructions in plain markdown files that we treat as the source of truth, and we treat any vendor-specific file as a build artifact we can regenerate. That decision costs almost nothing on day one and saves you the week you would otherwise lose when a vendor changes terms, pricing, or priorities.
Credit where due: Anthropic's response, working on hackability and easier AGENTS.md use rather than digging in, is the right direction. But a promise under pressure is not a shipped feature. We will believe the interoperability story when the file loads by default.
✅ What to do
Two concrete moves you can make today, whether you write code or just run your business through AI agents.
First, make your instructions portable. Keep one canonical instruction file for your agents and mirror it into whatever format each tool requires. On Mac or Linux, a symlink does this in one line, which is exactly the workaround engineers discussed in the Shopify thread. Second, run a quick lock-in audit: list every AI tool you depend on, and for each one write down what you would lose and how long migration would take if it changed direction tomorrow. If any answer scares you, that is your next weekend project.
- ✔Create one canonical AGENTS.md as your source of truth for agent instructions
- ✔Mirror it to CLAUDE.md with a symlink or a copy step so Claude Code stays in sync
- ✔List every AI tool your business depends on and what leaving it would cost
- ✔Prefer tools that read open formats when two options are otherwise equal
- ✔Watch Claude Code release notes for native AGENTS.md support before restructuring anything
🔄 The bigger pattern: standards wars come for AI agents
We have seen this movie before. Browsers fought over web standards, messaging apps fought over protocols, and cloud vendors fought over container formats. In each case the ecosystem eventually punished proprietary divergence, but only after users spent years paying the compatibility tax.
AI agent configuration is having that moment now. AGENTS.md for instructions, shared skills directories for capabilities, and the Model Context Protocol for tool connections are converging into a common stack. Vendors that fight the convergence will keep winning benchmarks and losing trust. The Shopify episode matters because it shows the market is now willing to price that trust: a CEO publicly treated a config file format as a dealbreaker for an entire company.
For buyers, that is good news. Every public fight like this one pushes the industry toward portability, and portability is the single best defense a small operator has against decisions made in rooms they will never sit in.
❓ Frequently Asked Questions
Did Shopify actually ban Claude Code?
No. As of late August 2026, CEO Tobi Lütke said he was considering a company-wide ban until Claude Code reads AGENTS.md and shared agent configuration standards. It was a public threat and negotiating position, not an enacted policy.
What is the difference between AGENTS.md and CLAUDE.md?
Both are markdown files that tell an AI coding agent how to behave in a project. AGENTS.md is an open convention read by tools like GitHub Copilot, Cursor, Codex, and Amp. CLAUDE.md is Anthropic's proprietary equivalent, and by default Claude Code reads only CLAUDE.md. The content is similar; the incompatibility is the problem.
Is Anthropic going to add AGENTS.md support to Claude Code?
Anthropic has signaled yes, but nothing has shipped as a native default. After the backlash, Claude Code team member Thariq Shihipar said the team is working on making Claude Code more hackable, including easier use of AGENTS.md and broader system-prompt customization. Until then, a symlink from CLAUDE.md to AGENTS.md is the common workaround.
Why does a config file matter to non-developers?
Instruction files are where you encode your business rules for AI agents. If they only work in one vendor's format, switching tools means rebuilding and re-testing everything. Open formats keep your setup portable, which keeps vendors competing for you instead of holding your configuration hostage.
🏁 Final Thoughts
The Claude Code AGENTS.md backlash is a small story about a file name and a big story about who controls your AI setup. The facts: Shopify's CEO threatened a ban on August 25, 2026, a roughly 5,000-upvote request had sat at "not planned" since May 2026, and public pressure got Anthropic to commit to easier AGENTS.md use. The lesson for solo operators: you cannot out-shout a CEO, so build portability into your agent stack now, starting with one canonical instruction file you own. Source: The New Stack, "Shopify's CEO threatened to ban Claude Code. Anthropic had already closed the feature request" (https://thenewstack.io/shopify-claude-code-agentsmd/), with syndicated coverage via 36Kr. Related reading on this blog: our earlier post on the MCP roadmap and where agent interoperability standards are heading. If you run real work through AI agents, subscribe for one field-tested breakdown like this every week, and tell us in the comments: would you ban a tool over a config file?
Last updated: August 27, 2026 · Keyword: Claude Code AGENTS.md · Agents at Work
Comments
Post a Comment