Keep agents busy, treat them as a pool rather than run them in batches

What variant of Codex are you using?

CLI

What feature would you like to see?

If Codex needs more agents than agents.max_threads, it runs them in “waves”: it starts up to agents.max_threads tasks, then waits for all of them to finish before starting the next wave.

I’d like to see agents.max_threads work like a worker pool with N slots: as soon as any agent finishes, if there are still pending tasks, start the next task immediately. Otherwise, a single slow task can hold up the entire next set of work.

Additional information

_No response_