The Pulse
Third Moment Runs 10,000 Agents for Up to 48 Hours
OpenAI says Third Moment Research runs 10,000 AI agents concurrently on investigations lasting 36 to 48 hours. The company describes its use of Codex, GPT-6 models and Daybreak for quantitative research, agent oversight and security testing
AI.info Team · Written with AI ·
Third Moment Research says it can run 10,000 AI agents at once on investigations lasting 36 to 48 hours, using OpenAI models and Codex to conduct and manage the work. OpenAI described the startup’s system on October 14, outlining applications in quantitative finance, automated experiments and security testing.
The scale is paired with a specific research loop: agents run experiments, assess results and choose follow-up work, while a separate harness can retry unfinished tasks or send them for human review. “In a single session, GPT-6.1 Sol is able to maintain a coherent thread of computational research across days of compute without losing the plot,” Adam Khakhar, co-founder of Third Moment Research, says in OpenAI’s account.
Codex cut a backtest from hours to minutes
Third Moment’s early work focuses on quantitative finance, where the team can measure whether a strategy performs better on historical data. OpenAI says the startup used Codex with GPT-5.6 to examine thousands of recorded agent trajectories, alongside timestamps and the research codebase. The analysis found that agents were wasting time and tokens checking whether a backtest had finished.
The test covered a year of order-level U.S. equity data. OpenAI says Codex identified a caching change and a memory-packing technique, then tried and deployed the fix. The company reports that the backtest time fell from about three hours to roughly three to five minutes, a 97% reduction listed in the post.
That improvement matters to the workflow because agents had been repeatedly polling for results while waiting. Shorter runs, OpenAI says, freed token capacity for additional experiments and reduced the amount of context taken up by status checks. The post does not provide a dollar-cost comparison or a third-party evaluation of the reported reduction.
A second harness handles unfinished tasks
For quantitative research, Third Moment runs the Codex harness in a Dockerized environment. Each agent receives a terabyte of data, documentation and a research repository containing tools for feature engineering, GPU-supported model training, trading strategies and simulation, according to OpenAI.
GPT-6 Astra, GPT-6 Sol and GPT-6.1 Sol supply the reasoning; Codex lets agents plan, act, set objectives and spawn subagents. When a task stops short of completion, an outer harness powered by GPT-6 Sol or GPT-6.1 Sol reviews the agent’s trajectory and decides whether to retry, end the job or refer it for human review. “The Codex harness is extremely versatile, and it enables use across many different domains,” Khakhar says.
The agents test strategies against historical data and use each result to decide what to try next. OpenAI says Third Moment’s work has improved the baseline trading system built by its engineers, but the post does not give a performance figure for that improvement.
Daybreak checks the agent workspaces
Running thousands of agents also creates a security problem: an agent should not escape its research environment and reach the wider network. Third Moment uses Daybreak, OpenAI’s cybersecurity platform, to test those environments before simulations begin. The checks are intended to find whether an agent can break out of its assigned workspace.
“Without Daybreak, we would not be able to autonomously test this at scale,” Khakhar says. OpenAI’s description explains the testing approach but gives no results from a specific penetration test or detail on how often a test found a vulnerability.
Third Moment’s goal is hundreds of thousands
The company says its longer-term plan is to move from individual investigations to hundreds of thousands of collaborating subagents, running chains of dozens of dependent experiments. OpenAI sketches possible business uses such as adjusting retail supply decisions when inventory runs short, finding performance bottlenecks in software and checking whether documents’ stated commitments match recorded outcomes.
For now, the concrete examples in the post center on quantitative research: speeding up a backtest, setting up agents with a large research environment and adding automated security checks. OpenAI presents the 10,000-agent runs and the 36-to-48-hour sessions as current capacity; the much larger network remains Third Moment’s stated plan.