Third Moment Research runs 10,000 simultaneous AI agents for stretches exceeding 36 hours using OpenAI's GPT-6 Astra and GPT-6.1 Sol models, with Codex handling code generation and Daybreak purpose-built for cybersecurity workloads. These are not chatbots answering questions. These are persistent, multi-day autonomous systems operating at a scale most organizations have not attempted.
The detail worth reading for: the agents improve with use, meaning performance compounds over the run duration rather than degrading. That is the operational claim that changes the calculus on long-horizon AI tasks. The piece also breaks down how Daybreak fits into security workflows specifically, which is rare technical transparency from an OpenAI blog post.
The open question this article raises but does not fully answer is failure mode management across 10,000 concurrent agents running for days. That gap is reason enough to read the original and form your own judgment on what Third Moment is actually claiming versus what has been validated.
[READ ORIGINAL →]