Third Moment Research runs 10,000 agents at once on OpenAI models

A self-improving system built around company-specific work

Third Moment Research starts from the idea that every company has its own workflows, tools, edge cases and definitions of success. In its view, the best agents should be strong at that specific work and get better the more they are used.

To get there, the company uses OpenAI's latest models to understand how agents behave, identify recurring weaknesses and automatically create thousands of synthetic scenarios around those problems. The team then uses the scenarios to improve the agents and repeats the process as new issues appear.

According to OpenAI, Codex, GPT-6 Astra, GPT-6 Sol and GPT-6.1 Sol support the research and computational work. The company also uses Daybreak to test the security of the environments where agents operate.

Cutting a backtest from hours to minutes

Quantitative finance was chosen as an early domain because success is easy to measure. It gives the team a concrete setting to study where agents struggle.

In one case, the team used Codex with GPT-5.6, the latest GPT model at the time, to analyze thousands of recorded simulation trajectories alongside timestamps and the codebase. Codex found a bottleneck: backtesting a trading strategy across a year of order-level US equity data took several hours. While waiting, agents kept polling for completion, which wasted tokens and filled their context with status checks.

After profiling the backtest and analyzing its inputs and usage, Codex identified a caching optimization and a memory-packing technique. It tested solutions and deployed the fix to production. The backtest dropped from approximately 3 hours to approximately 3 to 5 minutes, which reduced token use and context pollution and left more of the token budget for experiments.

How the agents are set up and kept on task

For quantitative research, the team runs the default Codex harness inside a Dockerized environment. Agents receive a terabyte of data, documentation and a research repository with modules for feature engineering, GPU-supported model training, trading strategies and simulation. The GPT-6 models provide the reasoning, while the harness lets agents act, create plans, set objectives, spawn subagents and compact context.

Some agents stop before finishing a task. To handle this, Third Moment treats Codex as a subagent inside an outer harness powered by GPT-6 Sol or GPT-6.1 Sol. When a task is left unfinished, the outer harness examines the trajectory and decides whether to retry, terminate the job or escalate to human review.

The agents work to improve a baseline trading system built by the company's engineers. They test strategies on historical data, judge whether an improvement is meaningful or noise, and use each result to pick the next experiment. According to co-founder Adam Khakhar, this work has improved the baseline system.

From prescribed experiments to unexpected findings

Khakhar says that before GPT-5.5, models could only run experiments prescribed by a user. With GPT-5.5 and GPT-5.6 in the Codex harness, models could use results to design follow-up experiments. With the GPT-6 models, agents can investigate root causes through hundreds of experiments running in parallel with subagents.

In one session, GPT-6.1 Sol identified the impact of non-displayed trades on an exchange where they made up an unexpectedly large share of volume. Khakhar called it "an unobvious result that our research team did not initially consider."

Third Moment tests frontier releases at their highest reasoning settings across optimization, machine learning systems and quantitative research. Testing now also includes GPT-6 Luna. Co-founder Jeffrey Yan says GPT-6.1 Sol is the most-used model because of its token efficiency: it "completes tasks with the fewest tokens while delivering performance similar to GPT-6 Astra."

On one benchmark, agents must cut the update latency of a live scoreboard service while preserving its outputs. Speedup is measured by dividing the average runtime of an expert-crafted solution by that of the model's solution on the same hardware, subject to correctness. According to Khakhar, GPT-6 Astra delivered the biggest speedup. For targeted work, such as checks for specific security weaknesses in the simulation system, the team uses more efficient models like GPT-5.6 Terra.

Security testing with Daybreak and plans for scale

With 10,000 agents running autonomously, Third Moment needs to test its research environments at scale. It uses the Daybreak cybersecurity platform to automate this before simulations launch. The tests check whether an agent could break out of its environment and reach the wider network. Khakhar says: "Without Daybreak, we would not be able to autonomously test this at scale."

Looking ahead, Third Moment plans to move from individual investigations to hundreds of thousands of collaborating subagents. Its roadmap calls for trillions of tokens supporting chains of dozens of dependent experiments, with the goal of autonomously improving agents as new challenges arise. The company points to uses such as supply chain decisions for retailers, performance bottlenecks in large codebases and mismatches between commitments and outcomes across documents. Khakhar expects people to spend more time thinking about how to do and optimize research rather than doing it themselves.