Third Moment Research is building a self-improving AI system that lets agents get better at specific company workflows the more they are used. Using OpenAI's latest models, the system identifies recurring agent weaknesses and automatically generates thousands of synthetic scenarios to address them. The company runs 10,000 agents at once on investigations lasting 36 to 48 hours, and uses Daybreak to test the security of the environments where those agents work.
One early focus is quantitative finance, where success is clearly measured. Using Codex with GPT-5.6, the team analyzed thousands of recorded simulation trajectories and found a bottleneck: backtesting a trading strategy across a year of order-level US equity data took several hours. Codex identified a caching optimization and a memory-packing technique, cutting the backtest from about 3 hours to about 3 to 5 minutes. This reduced token use and context pollution.
For quantitative research, each agent runs the Codex harness inside a Dockerized environment with a terabyte of data, documentation, and a research repository. GPT-6 Astra, GPT-6 Sol, and GPT-6.1 Sol provide reasoning, while an outer harness reviews unfinished tasks and decides whether to retry, terminate, or escalate to human review.
Co-founder Adam Khakhar says the models can now design follow-up experiments from their own results, rather than only executing prescribed ones. He credits the approach with improving the company's baseline trading system, and notes that GPT-6.1 Sol can keep a coherent thread of computational research across days of compute.
Source: OpenAI Blog · Summarized by HeadlinesBriefing