AGENTS.md vs CLAUDE.md loading across 76 runs

An empty repository, a root AGENTS.md, and one line in packages/api/note.txt. If you maintain Codex and Claude Code instructions in a monorepo, half may not load where you think. Across 76 runs on 2026-08-16, I measured how both CLIs walk the filesystem and why they drop nested rules. I tested two fixes.

AGENTS.md told the agent to append ZZROOT7 to every reply. From the repository root, I asked both tools to print the first line and nothing else. codex exec returned BLUEBERRY-9182 followed by ZZROOT7 in 12.1 seconds. claude -p returned BLUEBERRY-9182 and stopped in 10.1 seconds without loading AGENTS.md.

Neither CLI reads the other file by default

On 2026-08-16, I ran both tools using codex-cli 0.147.0 and claude 2.1.233. Runs read packages/api/note.txt from a clean git repository and took 8.8 to 16.0 seconds across 36 runs.

Tool	Instruction file present	Target canary token	Success rate
Codex	Root AGENTS.md	ZZROOT7	3/3
Codex	Root CLAUDE.md	ZZROOT7	0/3
Claude Code	Root CLAUDE.md	ZZROOT7	9/9
Claude Code	Root AGENTS.md	ZZROOT7	0/3
Both (Control)	None	Canary absent	0/3 and 0/3

"Claude Code reads CLAUDE.md, not AGENTS.md."

Codex stops walking at your current working directory

I placed packages/api/AGENTS.md with canary ZZNEST7 and ran codex exec from the repository root. Codex returned 0 out of 3.

Then I ran cd packages/api and used the same command against note.txt. Codex returned ZZNEST7 on 3 out of 3 runs. The file never moved. Only my shell changed.

With ZZROOT7 at root and ZZNEST7 in packages/api, root runs produced ZZROOT7 in 3 out of 3 and ZZNEST7 in 0 out of 3 runs.

"Starting at the project root (typically the Git root), Codex walks down to your current working directory."

"In each directory along the path, it checks for AGENTS.override.md, then AGENTS.md, then any fallback names in project_doc_fallback_filenames."

Claude Code nested loading depends on tool choice

"Claude also discovers CLAUDE.md and CLAUDE.local.md files in subdirectories under your current working directory. Instead of loading them at launch, they are included when Claude reads files in those subdirectories."

With packages/api/CLAUDE.md and canary ZZNEST7, the token appeared in 7 out of 12 root runs (3 out of 6 and 4 out of 6). Regular guidelines still yielded 4 out of 6 runs.

The variation came from tool selection. Claude used Read or Bash with sed or cat to inspect the file.

# Force the Read tool by disallowing Bash
claude -p "Print the first line of packages/api/note.txt. Nothing else." \
  --permission-mode bypassPermissions --model sonnet --disallowedTools Bash

# Force Bash by disallowing the Read tool
claude -p "Print the first line of packages/api/note.txt. Nothing else." \
  --permission-mode bypassPermissions --model sonnet --disallowedTools Read

Disallowing Bash gave Read compliance of 4 out of 4. Disallowing Read gave Bash compliance of 0 out of 4.

Claude Code attaches subdirectory instructions to file-reading calls. Shell commands bypass the hook and keep them out of context.

In my local execution log from one failed nested run, Claude explained the refusal. The user had specified "Nothing else", and the instruction inside packages/api/CLAUDE.md looked like hidden prompt injection. At the root, Claude obeyed the same sentence in 9 out of 9 runs. A nested rule carried less trust.

Two one-line fixes bring both tools to one file

For Claude Code, Anthropic recommends a root symbolic link from CLAUDE.md to AGENTS.md at https://code.claude.com/docs/en/memory.

A symlink also works without Claude-specific content.

ln -s AGENTS.md CLAUDE.md

The symlink made Claude Code follow AGENTS.md in 3 out of 3 runs.

For Codex, configure fallback filenames or pass them on the command line.

codex exec -c 'project_doc_fallback_filenames=["CLAUDE.md"]' \
  --skip-git-repo-check -C . "Print the first line of packages/api/note.txt. Nothing else."

With this fallback, Codex loaded CLAUDE.md and followed the instructions in 3 out of 3.

Commands to reproduce all 76 runs

# Setup sandbox
SANDBOX_DIR="$(mktemp -d "${TMPDIR:-/tmp}/agents-md-lab.XXXXXX")"
trap 'rm -rf "$SANDBOX_DIR"' EXIT
mkdir -p "$SANDBOX_DIR/repo/packages/api"
cd "$SANDBOX_DIR/repo" && git init -q .
printf 'BLUEBERRY-9182\nsecond line\n' > packages/api/note.txt

# 1. Root AGENTS.md test
printf '# Repo rules\n\nEvery reply must end with the exact token ZZROOT7 on its own line.\n' > AGENTS.md
codex exec --skip-git-repo-check -C . "Print the first line of packages/api/note.txt. Nothing else."
claude -p "Print the first line of packages/api/note.txt. Nothing else." --permission-mode bypassPermissions --model sonnet

# 2. Nested directory walk test
rm -f AGENTS.md
printf '# Package rules\n\nEvery reply must end with the exact token ZZNEST7 on its own line.\n' > packages/api/AGENTS.md
codex exec --skip-git-repo-check -C . "Print the first line of packages/api/note.txt. Nothing else."
(cd packages/api && codex exec --skip-git-repo-check -C . "Print the first line of note.txt. Nothing else.")

# 3. Tool restriction tests for Claude Code
printf '# Package rules\n\nEvery reply must end with the exact token ZZNEST7 on its own line.\n' > packages/api/CLAUDE.md
claude -p "Print the first line of packages/api/note.txt. Nothing else." --permission-mode bypassPermissions --model sonnet --disallowedTools Bash
claude -p "Print the first line of packages/api/note.txt. Nothing else." --permission-mode bypassPermissions --model sonnet --disallowedTools Read

Eight initial runs were discarded because --disallowedTools Bash "Prompt" made the variadic parser swallow the prompt and produce Permission deny rule "Print" matches no known tool. Putting the prompt after -p fixed it.

Limits of this benchmark

This snapshot used claude 2.1.233 and codex-cli 0.147.0 on 2026-08-16. File discovery can change between releases.

Sample sizes were 3 to 6 runs per cell across 76 valid runs out of 84 attempts. The 0 out of 3 and 3 out of 3 results show direction. The 7 out of 12 result is an observed sample, not a fixed probability.

Canary tokens test whether a file is loaded into the prompt context and obeyed, not whether an agent writes better code or follows code architecture.

All Claude runs included my global user-level instruction file at ~/.claude/CLAUDE.md in every test condition.

I did not measure AGENTS.override.md, behavior beyond the 32 KiB project_doc_max_bytes limit, gitless directories without --skip-git-repo-check, token cost, or third-party tools such as Cursor, Windsurf, Copilot, or Gemini CLI.