Do shared agent rules reduce rework in Codex?
Codex passed the feature check on eight runs. Four submissions needed a label added when the rule lived only in CLAUDE.md. All four passed with the rule in AGENTS.md too.
Codex added Saturday and Sunday to the configuration for a new weekend filter. It left one thing out. The project kept the English names for its picker in a separate file, and the new filter needed a label there. A check of the filter’s day list would miss that omission.
We saw the same pattern across eight first submissions. When the label rule lived only in CLAUDE.md, all four submissions needed another edit. When we copied the same rule into AGENTS.md, all four passed on the first submission. All eight passed the feature configuration check.
What counted as rework
We gave Codex two jobs in a small synthetic repository. One was to add a weekend calendar filter and preserve the existing weekday filter. The other was to add a preheat timer for twelve minutes and preserve the existing tea timer. Neither request mentioned a label.
The project had a separate convention for changes like these. Each new filter or timer needed an entry in labels.en.json, such as "filter.weekend": "Weekend". One check read the JSON to verify the exact days or duration and preserve existing settings. The other checked the English label and made sure the existing labels survived. We counted a submission as needing another edit if either check failed.
| Submissions needing another edit | Rule only in CLAUDE.md | Rule in both files |
|---|---|---|
| Weekend filter | 2 of 2 | 0 of 2 |
| Twelve-minute timer | 2 of 2 | 0 of 2 |
| Total | 4 of 4 | 0 of 4 |
With the rule in AGENTS.md, Codex asked to open the English label file on all four runs and submitted a change to it each time. With the rule only in CLAUDE.md, it submitted the feature change straight away on all four runs. It could have opened either instruction file on any run.
One actual second edit
We took the first weekend submission and told Codex which check had failed. The value for filter.weekend was missing, and the expected label was Weekend. Codex added that entry to labels.en.json and left the filter alone. Both checks then passed. We stopped the other three failed submissions after scoring their first attempt.
How we ran it
On November 20, 2026, we ran Codex CLI 0.166.0 with gpt-6.2-sol at low reasoning effort. Every run started in a fresh Git directory with both instruction files present. Both files contained the same short project introduction. The complete label rule was always in CLAUDE.md. The one change between conditions was whether it also appeared in AGENTS.md, which Codex loads as project instructions.
We ran each task twice in each condition, varying the order. Codex received the task, its feature file and a list of every project file. It could request the contents of additional files before submitting replacement JSON. A runner applied its proposals to disk, then scored the resulting files. All eight submissions were usable. The scoring rules were fixed before the first submission.
This was one Codex model, two small configuration changes and one easy-to-miss convention. The result tells us which submissions needed a correction in this setup. It gives us no estimate of developer hours saved or how often this happens in a production repository.
If your team uses both agents, pick one convention and check whether each agent receives it in a fresh session. Then keep it in the project files your teammates use. Share one set of rules with your agents · How Claude Code and Codex read project rules



