Shopify's Split Brain Is Not a Filename Dispute: It Is Drift, and One Standard File Will Not End It

untactit published a technical analysis of the AGENTS.md vs CLAUDE.md split brain issue raised by Shopify CEO Tobi Lutke, arguing the failure is instruction-file drift across copies and machines, with two free fixes.

On August 25, 2026, Shopify CEO Tobi Lutke posted that he was considering banning Claude Code at Shopify until it reads AGENTS.md and .agents/skills, saying that insisting on only reading CLAUDE.md sometimes leads to split brain problems when different team members use different tools. Source: https://x.com/tobi/status/2092259436538495186

A study published in February 2026 (arXiv 2602.14690, https://arxiv.org/html/2602.14690v1) analyzed 2,926 GitHub repositories using agentic coding tools. It found instruction and context files were the dominant configuration mechanism: CLAUDE.md appeared in 45.4 percent of repositories, AGENTS.md in 40.6 percent, and copilot-instructions.md in 35.1 percent. Those percentages sum well over 100 because many repositories carry two or three instruction files meant to say the same thing.

untactit, a control plane for AI-agent skills, memory and instruction files, published a technical analysis arguing that Shopify's failure is not a filename dispute. It is instruction-file drift across copies and machines. A rule is written once in one file. Someone copies it because a different tool reads a different filename, or a subdirectory or a laptop needs its own copy. Someone edits one copy: a build command changes, a rule is relaxed for one team, a path moves. Nothing tells anyone the copies have diverged, and the agent reading the stale copy does exactly what it was told, and it is wrong.

The analysis makes three points. First, split brain is drift, not a filename problem. Even a single standardized filename still gets copied across directories and machines, and silent divergence still happens inside that one filename. Second, standardizing on one filename such as AGENTS.md removes one multiplier on the problem, not the mechanism that causes it. Third, two practices are available today without any paid tooling: stop hand-editing generated copies of instruction files by generating every tool-specific file from one source and treating generated files as unwritable by humans, and detect when a copy has been touched anyway by hashing every context file and SKILL.md file across a repository, grouping the files that are supposed to be identical, and reporting which ones have diverged.

untactit published an open-source command-line tool, agent-drift, that performs this drift scan across a directory tree. It requires no service; it reads files and reports which copies of the same document have split. The tool is available at https://github.com/untactit/agent-drift and the full analysis is available at https://dev.to/untactit/shopifys-split-brain-is-not-a-filename-dispute-it-is-drift-and-one-standard-file-will-not-end-2ogo

About untactit: untactit is one library for skills, memory and instruction files. Changes are approved by a person, then pushed to every connected tool, including Claude Code, Cursor, GitHub Copilot, Codex and Gemini, without hand-editing copies. Drift detection reports which copy diverged. untactit is in early access and free to start at https://untactit.com

Press contact: untactit IR, ir@untactit.com.

Media Contact

untactit IR

untactit

15F Toranomon Hills Business Tower, 1-17-1 Toranomon, Minato-ku, Tokyo, Tokyo

03-3527-7891

https://untactit.com

Share this press release:

Have your own news to share? Submit Press Release Free