The best combination that Codex is currently using:

First, have it do a detailed, high-level discussion with GPT 5.6 on Sol High. Once a plan is reached, that’s not enough yet—also make it produce a complete checklist that passes human testing via the built-in browser.
Then, let it independently open another Sol High to implement it on the ground. On our side, we handle setting up the built-in browser to check the code according to the checklist, while also building the code for the subagent.
This single-subagent supervision mode is the most efficient and the most reliable.
After fully landing everything, then call the Codex security plugin to scan recent commits—right away, perfect.