Third Moment Research uses OpenAI models and Codex to run 10,000 agents concurrently on research sessions lasting 36–48 hours, while Daybreak tests the security of the environments where they work. The startup says its agents improve through repeated cycles: models analyze agent behavior, identify recurring weaknesses, and generate thousands of synthetic scenarios for further testing. The system supports research and computational work, including quantitative finance.

In one quantitative finance application, Codex analyzed thousands of recorded simulation trajectories, timestamps, and code to investigate a slow backtest. Running a trading strategy across a year of order-level US equity data took several hours, and agents repeatedly checked its status while waiting. Codex identified caching and memory-packing changes, then deployed a fix that cut the backtest from approximately 3 hours to approximately 3 to 5 minutes. OpenAI reports a 97% reduction in backtest time; the shorter run also reduced token use and status checks that occupied agents’ context.

For quantitative research, agents run the default Codex harness in a Dockerized environment with a terabyte of data, documentation, and a research repository containing feature engineering, GPU-supported model training, trading strategies, and simulation modules. GPT-6 Astra, GPT-6 Sol, and GPT-6.1 Sol provide reasoning, while the harness lets agents use tools, plan, set objectives, spawn subagents, and compact context. An outer harness powered by GPT-6 Sol or GPT-6.1 Sol reviews unfinished tasks and decides whether to retry, stop, or send them for human review. The agents test strategies against historical data and use results to choose further experiments; co-founder Adam Khakhar says their work has improved the team’s baseline trading system.

Khakhar says GPT-5.5 and GPT-5.6, used with Codex, let models design follow-up experiments based on results. With GPT-6 Astra, GPT-6 Sol, and GPT-6.1 Sol, agents can investigate root causes through hundreds of parallel experiments using subagents. In one session, GPT-6.1 Sol identified that non-displayed trades represented an unexpectedly large share of volume on an exchange, a factor the team had not initially considered. Third Moment tests models in optimization, machine learning systems, and quantitative research, including GPT-6 Luna. Co-founder Jeffrey Yan says GPT-6.1 Sol is the company’s most-used model because it uses the fewest tokens while performing similarly to GPT-6 Astra. In a latency benchmark requiring a live scoreboard service to preserve its outputs, GPT-6 Astra achieved the largest speedup, according to Khakhar.

Third Moment uses Daybreak to test whether agents can escape their research environments and access the wider network before simulations launch. The company says this automates penetration testing across its agent infrastructure. Its reported scale includes 10,000 concurrent agents, research sessions of 36–48 hours, and 1 trillion OpenAI tokens used in a single day.

The startup plans to expand from individual investigations to hundreds of thousands of collaborating subagents. Its roadmap calls for trillions of tokens to support chains of dozens of dependent experiments, with the aim of using research to improve agents as new challenges arise.

Third Moment Research uses OpenAI models and Codex to run 10,000 agents concurrently on research sessions lasting 36–48 hours, while Daybreak tests the security of the environments where they work. The startup says its agents improve through repeated cycles: models analyze agent behavior, identify recurring weaknesses, and generate thousands of synthetic scenarios for further testing. The system supports research and computational work, including quantitative finance.

In one quantitative finance application, Codex analyzed thousands of recorded simulation trajectories, timestamps, and code to investigate a slow backtest. Running a trading strategy across a year of order-level US equity data took several hours, and agents repeatedly checked its status while waiting. Codex identified caching and memory-packing changes, then deployed a fix that cut the backtest from approximately 3 hours to approximately 3 to 5 minutes. OpenAI reports a 97% reduction in backtest time; the shorter run also reduced token use and status checks that occupied agents’ context.

For quantitative research, agents run the default Codex harness in a Dockerized environment with a terabyte of data, documentation, and a research repository containing feature engineering, GPU-supported model training, trading strategies, and simulation modules. GPT-6 Astra, GPT-6 Sol, and GPT-6.1 Sol provide reasoning, while the harness lets agents use tools, plan, set objectives, spawn subagents, and compact context. An outer harness powered by GPT-6 Sol or GPT-6.1 Sol reviews unfinished tasks and decides whether to retry, stop, or send them for human review. The agents test strategies against historical data and use results to choose further experiments; co-founder Adam Khakhar says their work has improved the team’s baseline trading system.

Khakhar says GPT-5.5 and GPT-5.6, used with Codex, let models design follow-up experiments based on results. With GPT-6 Astra, GPT-6 Sol, and GPT-6.1 Sol, agents can investigate root causes through hundreds of parallel experiments using subagents. In one session, GPT-6.1 Sol identified that non-displayed trades represented an unexpectedly large share of volume on an exchange, a factor the team had not initially considered. Third Moment tests models in optimization, machine learning systems, and quantitative research, including GPT-6 Luna. Co-founder Jeffrey Yan says GPT-6.1 Sol is the company’s most-used model because it uses the fewest tokens while performing similarly to GPT-6 Astra. In a latency benchmark requiring a live scoreboard service to preserve its outputs, GPT-6 Astra achieved the largest speedup, according to Khakhar.

Third Moment uses Daybreak to test whether agents can escape their research environments and access the wider network before simulations launch. The company says this automates penetration testing across its agent infrastructure. Its reported scale includes 10,000 concurrent agents, research sessions of 36–48 hours, and 1 trillion OpenAI tokens used in a single day.

The startup plans to expand from individual investigations to hundreds of thousands of collaborating subagents. Its roadmap calls for trillions of tokens to support chains of dozens of dependent experiments, with the aim of using research to improve agents as new challenges arise.