--max-concurrent gates episodes, so a multi-agent env's internal fan-out multiplies the number of live agent rollouts — and with them runtime boxes and in-flight completions — invisibly (e.g. -c 128 --env.id best-of-n --env.n 16 allows 2048 concurrent agent runs). We need a configurable bound on practical agent concurrency (global cap on live Agent.runs, possibly per-role overrides), rather than only the logical episode count.
--max-concurrentgates episodes, so a multi-agent env's internal fan-out multiplies the number of live agent rollouts — and with them runtime boxes and in-flight completions — invisibly (e.g.-c 128 --env.id best-of-n --env.n 16allows 2048 concurrent agent runs). We need a configurable bound on practical agent concurrency (global cap on liveAgent.runs, possibly per-role overrides), rather than only the logical episode count.