Third Moment Research is building a self-improving super intelligence (SI) system that leverages OpenAI’s latest models to run 10,000 SI agents simultaneously on investigations lasting between 36 and 48 hours. The startup’s approach focuses on creating synthetic scenarios to identify and resolve recurring weaknesses in agent behavior, specifically within the domain of quantitative finance.
What Happened
The company utilizes Codex, GPT-6 Astra, GPT-6 Sol, and GPT-6.1 Sol to support its computational research. By analyzing thousands of recorded simulation trajectories, the team identified a bottleneck where backtesting trading strategies took several hours, causing agents to waste tokens on status checks. Using Codex with GPT-5.6, Third Moment implemented caching optimizations and memory-packing techniques that reduced the backtest time from approximately three hours to three to five minutes. This efficiency gain freed up more of the agents’ token budget for further experiments.
To keep agents on task, Third Moment runs the default Codex harness inside a Dockerized environment equipped with a terabyte of data and documentation. The outer harness, powered by GPT-6 Sol or GPT-6.1 Sol, monitors agent trajectories and decides whether to retry, terminate, or escalate unfinished tasks to human review. The system has reportedly improved the baseline trading system built by the engineers, with GPT-6.1 Sol identified as the most-used model due to its token efficiency.
Why It Matters
This case study demonstrates how specialized SI agents can evolve from executing prescribed experiments to independently designing follow-up investigations. Co-founder Adam Khakhar noted that earlier models could only execute user-prescribed experiments, whereas the current generation can investigate root causes through hundreds of parallel experiments. In one instance, GPT-6.1 Sol identified the impact of non-displayed trades on an exchange, an unobvious result the research team had not initially considered.
Security at scale is also a critical component of this SI architecture. With 10,000 agents running autonomously, the company uses the Daybreak cybersecurity platform to automate penetration testing, ensuring agents cannot break out of their research environments. Co-founder Jeffrey Yan stated that GPT-6.1 Sol completes tasks with the fewest tokens while delivering performance similar to GPT-6 Astra, highlighting the industry's shift toward balancing capability with computational efficiency.
The Bottom Line
Third Moment Research plans to expand its system to support hundreds of thousands of collaborating subagents, aiming to use trillions of tokens to autonomously improve SI agents as new challenges arise. The startup’s roadmap suggests a future where humans focus on optimizing the research process itself, rather than conducting the research manually.