The coolest combination Codex is using right now: first, have it discuss with GPT 5.6 sol high in depth; after coming up with a solution, it’s not enough—you also need to have it produce a complete checklist that passes manual testing inside the built-in browser. Then let it open another sol high on its own to implement the solution on the ground. In parallel, you handle opening the built-in browser to check the code according to the checklist, and also supervise the code of the subagent. This kind of single-subagent oversight mode is the most efficient and the most reliable. After it’s fully implemented, then call the codex security plugin to scan recent commits—straight to a perfect result.