Third Moment Research runs 10,000 agents for 36+ hours with GPT-6

The lab applies this setup to quantitative finance, where agents test trading strategies on historical data. In one instance, Codex identified a bottleneck in backtesting US equity data, reducing processing time from approximately 3 hours to 3 to 5 minutes through caching and memory-packing optimizations. This efficiency freed up token budgets for further experiments. The team notes that newer models like GPT-6.1 Sol maintain coherent research threads over days of compute, enabling agents to investigate root causes through hundreds of parallel experiments.

Security is managed using OpenAI’s Daybreak platform, which automates penetration testing to ensure agents cannot break out of their Dockerized environments. Third Moment plans to scale this architecture to hundreds of thousands of collaborating subagents, aiming to let humans focus on optimizing the research process rather than executing individual tasks.