Third Moment Research is building a self-improving AI system around the idea that agents should specialize in a company’s specific workflows and improve with use. It uses OpenAI’s latest models to understand agent behavior, identify recurring weaknesses, and automatically generate thousands of synthetic scenarios around those problems; the team then uses those scenarios to improve agents and repeats the process as new issues emerge. Codex, GPT-6 Astra, GPT-6 Sol, and GPT-6.1 Sol support the research, allowing the company to run 10,000 agents at once on investigations lasting 36-48 hours. Daybreak is used to test the security of the environments where those agents work.

One early application is quantitative finance because it offers a clear measure of success. The team uses Codex to find and address problems in the research environment. In one instance, Codex with GPT-5.6 analyzed thousands of recorded simulation trajectories, timestamps, and the codebase. It uncovered a bottleneck: backtesting a trading strategy across a year of order-level US equity data took several hours, and agents waiting for completion repeatedly polled, wasting tokens and filling context with status checks. After profiling the backtest and analyzing inputs and usage, Codex identified a caching optimization and a memory-packing technique. It deployed the fix, reducing the backtest from roughly 3 hours to roughly 3-5 minutes, which reduced token waste and context pollution.

The agents run inside the default Codex harness in a Dockerized environment, with a terabyte of data, documentation, and a research repository containing modules for feature engineering, GPU-supported model training, trading strategies, and simulation. GPT-6 Astra, GPT-6 Sol, and GPT-6.1 Sol provide the reasoning; the Codex harness gives agents tools to act, create plans, set objectives, spawn subagents, and compact context. To handle tasks left unfinished, Third Moment treats Codex as a subagent within an outer harness powered by GPT-6 Sol or GPT-6.1 Sol. That outer harness inspects the trajectory and decides whether to retry, terminate, or escalate to human review. According to Third Moment, the agents’ work has improved the baseline trading system built by its engineers.

Third Moment describes a model progression underlying this. Before GPT-5.5, models could only execute experiments prescribed by a user. With GPT-5.5 and GPT-5.6 and the Codex harness, models could use results to design follow-up experiments. Now GPT-6 Astra, GPT-6 Sol, and GPT-6.1 Sol allow agents to investigate root causes through hundreds of parallel experiments with subagents. In one session, GPT-6.1 Sol identified the impact of non-displayed trades on an exchange where those trades accounted for an unexpectedly large share of volume, an unobvious result the research team had not initially considered.

The company evaluates models on optimization, machine learning systems, and quantitative research, testing frontier releases at highest reasoning settings for repository navigation, code efficiency, and sustained investigation. Current testing includes GPT-6 Luna. Co-founder Jeffrey Yan says GPT-6.1 Sol is the most-used model because of token efficiency: it completes tasks with the fewest tokens while delivering performance similar to GPT-6 Astra. In a benchmark asking agents to reduce a live scoreboard service’s update latency while preserving outputs, the team measures speedup by dividing an expert-crafted solution’s average runtime by the model’s solution runtime on the same hardware, subject to correctness; Adam says GPT-6 Astra delivered the biggest speedup. For targeted checks, including specific security weaknesses in the simulation system, the team uses more efficient models like GPT-5.6 Terra, while deep research can justify higher cost given the value of finding a signal in quantitative finance.

Daybreak is used to automate security testing at scale before simulations launch. The tests check whether an agent could break out of its research environment and access the wider network. Adam says that without Daybreak, they would not be able to autonomously test at this scale.

Third Moment’s broader ambition covers specialized daily work: supply chain decisions, performance bottlenecks in large codebases, and document commitments versus outcomes. Building on GPT-6 Astra, GPT-6 Sol, and GPT-6.1 Sol, it plans to scale from individual investigations to hundreds of thousands of collaborating subagents, with trillions of tokens supporting chains of dozens of dependent experiments, and to use that research to autonomously improve agents as new challenges arise. The stated goal is to put humans in a more leveraged position, thinking about how to optimize research itself rather than doing the research directly.