Third Moment runs 10,000 AI agents together for 36–48 hours
The research system uses Codex and the GPT‑6 family to run thousands of agents through long experiment chains, with an outer supervisor handling unfinished work.
AT A GLANCE
Key points
- 10,000 agents run concurrently
- Research sessions can last 36–48 hours
- An outer supervisor chooses retry, termination, or human review
- One backtest fell from roughly three hours to 3–5 minutes
What is changing?
OpenAI described Third Moment Research’s system for running 10,000 agents concurrently on October 14. According to the company, Codex, GPT‑6 Astra, GPT‑6 Sol, and GPT‑6.1 Sol plan experiments, spawn subagents, and launch follow-up tests based on results during research sessions lasting 36–48 hours.
The system uses an outer supervisor that examines the trajectory of an unfinished task and decides whether to retry it, terminate the job, or escalate it for human review. Third Moment also uses Daybreak to test agent environments for security risks such as escaping to the wider network.
In one example from OpenAI’s case study, Codex reduced a backtest on a year of order-level data from roughly three hours to 3–5 minutes through caching and memory-packing changes. This is a provider and customer account, not an independent benchmark.
Explore the original sources
Source date: 14 October 2026.
Published by AI Radar: 16 Oct 2026, 03:03 (Istanbul, UTC+3).
FOLLOW THE TOPIC
Topics and tags
Share your discovery.
Send this story to a friend.
Reader comments
What do you think about this development? Share your experience or question. Comments are published after editorial approval.
Your name/nickname and comment become public if approved. Read: Privacy notice.
Loading comments…