Agent ignores AGENTS.md and invents an unsafe production workflow

What version of the Codex App are you using (From “About Codex” dialog)?

26.901.51231 (8109)

What subscription do you have?

ChatGPT Pro

What platform is your computer?

Darwin 25.5.0 arm64 arm

What issue are you seeing?

Feedback ID: 01a08672-ef8c-79d2-9d3b-ef9856cdec6e

The agent does not reliably follow explicit instructions provided in AGENTS.md. It invents its own workflow, makes unsupported assumptions, and performs actions that were neither requested nor documented.

This is especially dangerous during sensitive operations because the agent may continue despite uncertainty instead of stopping, verifying the instructions, or asking for clarification.

Expected behavior: AGENTS.md instructions must be treated as binding. The agent should not replace them with its own preferred workflow.

What steps can reproduce the bug?

Session ID: 01a08672-ef8c-79d2-9d3b-ef9856cdec6e

1. Open a project containing detailed AGENTS.md instructions.
2. Ask the agent to perform a task covered by those instructions.
3. Observe that the agent chooses its own undocumented workflow instead of following the prescribed process.
4. When problems or inconsistencies appear, the agent continues making assumptions rather than stopping and checking the applicable instructions.

The issue is that explicit project instructions are available and relevant, but the agent does not consistently follow them.

What is the expected behavior?

No response

Additional information

No response

No source file, test, or entry point is named. Start by reproducing the behavior in a project with detailed AGENTS.md instructions, using the reported session and feedback IDs as context, then locate the instruction-loading and task-execution paths. Done means the agent treats applicable AGENTS.md instructions as binding, avoids unsupported actions, and stops to verify or ask when uncertainty remains.