In an October 14 OpenAI case study, Third Moment Research describes running up to 10,000 agents concurrently on investigations lasting 36–48 hours. Its stack combines GPT models, Codex for coding work and Daybreak to test environment security.
A backtesting example falls from roughly three hours to three to five minutes in the vendor/customer account. This is not a controlled benchmark or a promise that any application supports that concurrency or duration; it concerns one organization and its workflows.