Third Moment Research runs 10,000 AI agents for up to 48 hours in quantitative research

OpenAI published a case study on Third Moment Research, a startup that uses the company's models to build an agent system that tries to improve itself through use. The core idea: instead of a generic agent, each agent should get better and better at a specific task, with the system itself identifying where it fails and generating synthetic scenarios to fix that.

The testing ground the company chose is quantitative financial market research. Not by accident: it's a domain with an objective success criterion (the backtest returned more or less than expected), which makes it easier to assess whether an agent actually improved or just seemed to. According to OpenAI, Third Moment already runs 10,000 simultaneous agents in investigations that last 36 to 48 hours, using Codex and the GPT-6 Astra, GPT-6 Sol, and GPT-6.1 Sol models.

How the agents stay active for two days without losing the thread

Keeping an agent coherent for 36 hours of computation is a context-engineering problem, not just a matter of model capability. Third Moment solves this with a two-layer architecture: the standard Codex harness runs inside a Dockerized environment, with access to a terabyte of data, documentation, and a research repository with modules for feature engineering, GPU model training, trading strategies, and simulation.

On top of that, the company treats Codex itself as a subagent inside an external harness controlled by GPT-6 Sol or GPT-6.1 Sol. When a task is left unfinished, this external harness examines the execution trajectory and decides between three paths: retry, terminate the job, or escalate for human review. It's this supervisor that prevents the classic scenario of a stuck agent repeating the same action without making progress.

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

The backtest that dropped from 3 hours to 3-5 minutes

The most concrete example in the case study shows Codex analyzing thousands of logged simulation trajectories (the agents' actions and the environment's responses), along with timestamps and source code. Using GPT-5.6, which was OpenAI's most recent model at the time, Codex found a bottleneck: running the backtest of a trading strategy against a year of order-by-order US stock data took several hours.

While waiting, each agent kept repeatedly checking whether the backtest had finished, burning tokens and cluttering its own context with status checks. After analyzing the backtest's execution profile and usage patterns, Codex identified a cache optimization and a memory-packing technique, tested the solutions, and pushed the fix to production. The result: the backtest dropped from roughly 3 hours to 3 to 5 minutes, freeing up token budget for more experiments.

That's the kind of gain that matters to anyone operating agents in production: it's not the model getting "smarter" in the abstract, it's tokens wasted on polling turning into real research capacity. It's worth noting that the case study doesn't disclose the computational cost of the analysis Codex performed to arrive at this optimization, nor how many attempts it took to get it right.

Which model for which task: the cost lesson the case leaves between the lines

One of the most useful points in the material is how Third Moment chooses a model by type of work, rather than using a single model for everything. According to Jeffrey Yan, co-founder of the company, GPT-6.1 Sol is today the most used model internally precisely because of its token efficiency.

GPT-6.1 Sol is currently our most-used model given its token efficiency. It completes tasks with the fewest tokens while delivering performance similar to GPT-6 Astra.

Jeffrey Yan, co-founder of Third Moment Research

The division of tasks by model follows a cost-benefit logic that any team running agents at scale should recognize:

For deep research, Adam Khakhar argues it's worth paying more for stronger reasoning: according to him, in the quantitative research space, the value of finding an interesting signal is extremely high, which justifies the extra cost of running the more expensive model when the bet pays off.

Testing its own cage: security with Daybreak

Running 10,000 autonomous agents at the same time raises an obvious question: what stops an agent from escaping the simulation environment and accessing the company's broader network? Third Moment's answer is to use Daybreak, OpenAI's own cybersecurity platform, to automate penetration testing on these environments before each simulation is launched.

In practice, the agents themselves pentest the infrastructure they're about to run on, before starting the research work. This solves a real scale problem: manually testing the security of thousands of isolated environments, one by one, isn't feasible.

By using Daybreak, we're able to have agents themselves do penetration testing on the simulation and test the infrastructure and security of it. Without Daybreak, we would not be able to autonomously test this at scale.

Adam Khakhar, co-founder of Third Moment Research

What the case doesn't show

It's worth separating what's a replicable lesson from what's well-told marketing. The material comes directly from OpenAI, about a client that uses almost the entire OpenAI product portfolio (Codex, GPT-6 Astra, GPT-6 Sol, GPT-6.1 Sol, and Daybreak), so the numbers chosen tend to be the most favorable ones.

There is, for instance, no figure for the computational cost of keeping 10,000 agents active for 36 to 48 hours, nor a failure rate for investigations that end up escalated to human review instead of being completed by the agent. The gain from "3 hours to 3 to 5 minutes" is real and specific, but it's an isolated case, not an average.

It's also worth remembering that quantitative research is a particularly favorable domain for this kind of architecture: the success of a trading strategy is directly measurable in objective numbers. In more ambiguous domains (customer service, legal analysis, content moderation), defining what counts as "improvement" to feed the self-improvement cycle is much harder, and the case study doesn't address that.

What can be taken from this material, even outside fintech, are three concrete architecture patterns: an external harness supervising subagents with an explicit retry/terminate/escalate policy, attention to tokens wasted on polling as a real source of cost, and automated security testing of the sandbox itself before releasing autonomous execution at scale.

Next step: hundreds of thousands of agents collaborating

Third Moment describes a roadmap expanding from individual investigations to hundreds of thousands of collaborating subagents, with trillions of tokens sustaining chains of dozens of interdependent experiments. The stated goal is to use this research so that the agents themselves improve as new challenges arise, without repeating today's manual cycle.

For Adam Khakhar, this shift changes the role of people who work in research: people would start thinking about how to do research and optimize the research process itself, rather than directly executing the research. It's a big bet, still without public metrics on how close this is to working outside Third Moment's controlled lab.