Before you add more to AGENTS.md, check whether plain Codex already beats your setup

#What changed

On Oct 5, developer dax (@thdxr) made a point that drew about 1.9k likes: many custom workflows and setups patch problems that newer models no longer have, so someone using plain Codex may be getting closer to the best current behavior than a heavy tinkerer. Replies echoed it, with several people saying their long AGENTS.md files had turned into lists of workarounds for bugs that were fixed months ago.

Jarad Johnson (@jdjohnson) gave a concrete case. He says Codex worked noticeably better on a freshly set-up Mac Mini than on his daily machine, and he went back to short AGENTS.md files that mostly tell the model where to find context.

#Who is affected

Codex users running locally (CLI and IDE extension) who keep a global ~/.codex/AGENTS.md, long repo-level AGENTS.md files, or custom skills and wrappers written for older models. Teams that rely on shared quality rules are less affected; Johnson says organizations still need systems around the model.

#What to do now

1. See what Codex actually loads. From your repo, run codex --ask-for-approval never "List the instruction sources you loaded." Codex reads the global file in ~/.codex first, then at most one AGENTS.md per directory from the repo root down to your working directory, and stops adding files once the combined size reaches 32 KiB by default (project_doc_max_bytes).
2. Try a lean global setup without deleting anything. Put a short ~/.codex/AGENTS.override.md in place; Codex uses it instead of ~/.codex/AGENTS.md. Delete the override to go back.
3. Run the same real task both ways and compare the result, how many corrections you had to make, and how long it took.
4. Cut lines that work around old model behavior. Keep pointers to where things live (docs, scripts, test commands) and hard rules such as approved dependencies or files not to touch.
5. Repeat the check after each major model update.

#What is not confirmed

The idea that stock Codex beats tuned setups is an opinion plus individual reports, not a benchmark, and it has not been independently verified. Results will vary by repo and task. The commands and file-loading rules above come from OpenAI's Codex AGENTS.md guide, checked Oct 5, 2026.