Third Moment Research is building a self‑improving AI system based on the idea that the most useful agents should excel at a company’s specific work and get better the more they are used. The lab uses OpenAI models — notably Codex, GPT‑6 Astra, GPT‑6 Sol and GPT‑6.1 Sol — to analyze agent behavior, uncover recurring weaknesses, and automatically generate thousands of synthetic scenarios around those problems.
Scale and operational model
The company’s infrastructure can run 10,000 autonomous agents concurrently, with individual investigations lasting 36–48 hours. Models provide reasoning and oversight while Codex helps find and fix problems in the research environment. Identified issues become synthetic test cases which agents use to improve; the cycle repeats as new issues appear.
For security at scale, Third Moment uses the Daybreak cybersecurity platform to automate tests of the research environments before simulations run, checking whether an agent could break out of its sandbox and access wider network resources.
A concrete outcome from quantitative finance
An early application area is quantitative finance, because success is easy to measure. Third Moment’s agents work on improving a baseline trading system: they test strategies on historical data, judge whether changes are meaningful or noise, and use each outcome to choose the next experiment.
Using Codex together with earlier GPT models (including GPT‑5.6), the team analyzed thousands of recorded simulation trajectories — agents’ actions, environment responses, timestamps and the codebase. Codex identified a bottleneck: backtesting a trading strategy across a year of order‑level US equity data took several hours, and during that time agents repeatedly polled for completion, wasting tokens and filling the working context with status checks.
After profiling the backtest and its inputs, Codex suggested a caching optimization and a memory‑packing technique. The experiment and subsequent production deployment reduced the backtest runtime from about 3 hours to roughly 3–5 minutes. That cut token usage and context pollution, freeing more of each agent’s token budget for further experiments.
How agents are provisioned and coordinated
Third Moment gives each agent the tools, information and environment needed for the specific task. For quantitative research the team runs the default Codex harness in a Dockerized environment: agents receive a terabyte of data, documentation, and a research repository with modules for feature engineering, GPU‑backed model training, trading strategies and simulation.
GPT‑6 Astra, GPT‑6 Sol and GPT‑6.1 Sol provide reasoning, while the Codex harness gives agents the ability to act: create plans, set objectives, spawn subagents and compact context as work proceeds. When agents stop before completing a task, Third Moment treats Codex as a subagent inside an outer harness powered by GPT‑6 Sol or GPT‑6.1 Sol; the outer layer inspects the trajectory and decides whether to retry, terminate, or escalate to human review.
Model evaluation and cost tradeoffs
Third Moment evaluates models across optimization, machine learning systems and quantitative research dimensions. They test frontier releases at high reasoning settings to assess repository navigation, efficient code generation and sustained investigation capability. Current tests also include GPT‑6 Luna.
One benchmark asks agents to reduce a live scoreboard service’s update latency while preserving outputs; speedup is measured by dividing the expert solution’s average runtime by the model’s runtime on the same hardware, provided correctness is maintained. According to Adam, GPT‑6 Astra produced the largest speedup in that benchmark.
For targeted, lower‑cost tasks — for instance checks for specific security weaknesses in the simulation system — the team uses more efficient models like GPT‑5.6 Terra. For deep research, Adam argues stronger reasoning can justify higher cost: “In the space of quantitative finance, the value of finding some interesting signal is extremely high.”
Roadmap: many more collaborating subagents and trillions of tokens
Third Moment plans to expand from individual investigations to hundreds of thousands of collaborating subagents. Their roadmap envisions chains of dozens of dependent experiments supported by trillions of tokens, with the goal that ongoing research autonomously improves agents as new challenges arise.
Adam says this approach would let humans take a more leveraged role in shaping how agents learn and work: “Humans can be in a much more leveraged position, where they can think about how to do research and optimize the research itself instead of actually doing the research.”
The Third Moment methodology illustrates an approach to building specialized, scalable agent fleets tailored to company workflows, while simultaneously addressing security testing and cost‑benefit decisions.
