Third Moment Research runs 10,000 agents for 36+ hours using GPT-6 and Codex

Third Moment Research is building a self-improving AI system that runs 10,000 agents concurrently on investigations lasting 36–48 hours. The startup leverages GPT-6 Astra, GPT-6 Sol, and GPT-6.1 Sol alongside Codex to understand agent behavior, identify weaknesses, and automatically generate synthetic scenarios for improvement. This approach allows the system to maintain coherent computational research threads across days of compute without losing context.

Optimizing Quantitative Finance Workflows

The team applies this system to quantitative finance, where success is clearly measurable. By analyzing simulation trajectories with Codex and GPT-5.6, they identified a bottleneck in backtesting trading strategies that took several hours. Codex implemented a caching optimization and memory-packing technique, reducing backtest time from approximately 3 hours to 3 to 5 minutes. This 97% reduction in runtime significantly lowered token waste and context pollution, freeing up budget for further experiments.

Sustaining Long-Horizon Research

Agents operate within a Dockerized environment equipped with a terabyte of data and specialized modules for feature engineering and model training. The Codex harness enables agents to create plans, spawn subagents, and compact context, while GPT-6.1 Sol is noted for its token efficiency, completing tasks with the fewest tokens while matching the performance of GPT-6 Astra. This progression from GPT-5.5 to GPT-6 models allows agents to move from executing prescribed experiments to investigating root causes through hundreds of parallel subagent experiments.

Security and Future Scaling

To secure these autonomous environments, Third Moment uses the Daybreak cybersecurity platform to automate penetration testing, ensuring agents cannot break out of their sandbox. The company’s roadmap aims to expand from individual investigations to hundreds of thousands of collaborating subagents, supporting trillions of tokens in chains of dependent experiments. This shift is designed to place humans in a more leveraged position, focusing on optimizing research methodology rather than performing the research itself.