發生了什麼事
The New Stack reports that Shopify CEO Tobi Lütke threatened to ban Claude Code unless Anthropic adds native support for AGENTS.md and related agent skill files. The dispute concerns how coding agents receive persistent repository instructions, not Claude Code’s general coding ability.
The New Stack reports that Shopify CEO Tobi Lütke said on X that he was considering banning Claude Code at Shopify until Anthropic changed how the tool handles repository instruction files. Specifically, Lütke objected to Claude Code’s lack of native support for AGENTS.md and related .agents/skills files. The article presents the dispute as a compatibility and workflow problem involving AI coding agents, rather than as a judgment that Claude Code produces poor code.
According to The New Stack, Shopify has thousands of developers working in a large monorepo, and those developers do not all use the same AI coding tools. Repository instruction files can specify build commands, testing requirements, coding conventions, and other rules that an agent should follow. Because these files may exist at different levels of a directory tree, an agent can receive guidance based on the part of the repository it is handling.
The article says Claude Code uses CLAUDE.md files for project instructions and can discover relevant files in different parts of a repository, but does not read AGENTS.md natively. The New Stack reports that Anthropic offers workarounds, including importing AGENTS.md from CLAUDE.md or using symbolic links. Those approaches require teams to keep multiple representations of instructions synchronized, which can create additional maintenance work in a large codebase.
The New Stack also reports that requests for native AGENTS.md support have continued for nearly a year. It identifies a more recent request concerning recursive AGENTS.md discovery that Anthropic marked as “not planned.” The reported statements, -request status, and Shopify’s internal automation were not independently confirmed from primary documentation in the supplied material.
為什麼這很重要
In large monorepos, different agents may follow different build, testing, and coding instructions. The New Stack says Shopify has built automation to synchronize these instructions, but Lütke considers that an avoidable complexity cost for platform teams.
The practical issue is that AI coding agents depend on local context. If one agent reads a repository’s applicable instructions and another does not, the two tools may make changes under different assumptions about testing, formatting, architecture, or permitted workflows. The New Stack reports that Lütke described this as a “complexity tax” for engineering teams, because Shopify has to compensate for differences among tools.
This matters most in organizations where several agents work against the same codebase. A single developer can potentially maintain a simple workaround, but a platform team supporting thousands of developers must distribute, update, and verify those workarounds across many directories and tools. The New Stack’s account suggests that the cost is organizational as much as technical: engineers responsible for internal infrastructure become responsible for translating or synchronizing instructions that vendors do not interpret consistently.
The disagreement also illustrates a standardization problem in the AI developer-tool market. The New Stack reports that OpenAI introduced AGENTS.md in August 2025 for project-specific coding-agent instructions, and that OpenAI later said the format was used by more than 60,000 open-source projects and agent frameworks, with support from tools including Codex, Cursor, Gemini CLI, GitHub Copilot, Jules, and Visual Studio Code. The article says the format was later transferred to the Agentic AI Foundation under the Linux Foundation. These adoption figures and governance details come from The New Stack’s report and were not independently verified here.
The issue is not proof that one instruction-file format is universally superior, nor does it establish that Claude Code is less capable at software development. It shows that the surrounding control layer—how an agent receives and applies repository rules—can affect enterprise usability. The New Stack also discloses that its owner, Insight Partners, is an investor in OpenAI and Anthropic, a relevant consideration when assessing the outlet’s independence on a story involving both companies.
互動機制：它實際上是如何運作的
以互動方式探索這項發展背後的基礎技術。
crm_get_transaction(id='4092').An agent must create a draft calendar event for Tuesday at 2 p.m. Which evidence would establish the requested result?
接下來看什麼
Watch whether Anthropic changes Claude Code’s handling of AGENTS.md, whether other large engineering organizations raise the same concern, and whether repository instruction formats converge across coding-agent vendors. The New Stack’s account and the reported posts were not independently confirmed here.
The immediate question is whether Anthropic will add native AGENTS.md discovery or provide another mechanism that avoids duplicate files and manual synchronization. The New Stack reports that the relevant request was closed as “not planned,” but the supplied material does not establish whether that decision remains current or whether Shopify’s criticism has changed Anthropic’s roadmap.
Engineering organizations will also be watching for broader convergence among coding-agent tools. If AGENTS.md becomes a widely supported convention, teams may reduce the number of vendor-specific instruction files they maintain. If vendors continue to favor separate formats, platform teams may need translation layers, repository automation, symlinks, or policy controls to keep agents aligned. The article does not provide comparative measurements of the cost, reliability, or failure rate of these approaches.
A further issue is verification. The New Stack reports that Shopify already uses automation to address the mismatch, but the article does not describe that automation in enough detail to assess how well it works, how often synchronization fails, or whether it changes the behavior of agents in production. It also does not provide an on-record response from Anthropic to Lütke’s threat or independently document the request’s history.
The broader practical test will be whether instruction files remain reliable as repositories become more complex and more agents operate within them. Useful follow-up reporting would establish which tools support recursive discovery, how precedence is determined when files conflict, whether developers can audit the instructions an agent actually applied, and whether major customers treat format compatibility as a procurement requirement. No such outcomes are established by this report.