Third Moment Research runs 10,000 agents at once
AI lab Third Moment Research runs 10,000 concurrent agents through 36–48-hour research sessions on OpenAI models and Codex, giving a rare concrete look at how long-horizon agent harnesses work in production.
Third Moment Research is building a self-improving agent system. It studies how its agents behave, finds weaknesses that keep recurring, and generates thousands of synthetic scenarios to train those weaknesses out, OpenAI said in a customer case study. Quantitative finance is the lab's first test bed because success there has a clear measure.
The lab runs the default Codex harness inside Docker. Each agent gets a terabyte of data plus a research repository covering feature engineering, GPU model training, trading strategies and simulation. Agents sometimes stop before a task is done. To handle that, the lab treats Codex as a subagent inside an outer harness powered by GPT‑6 Sol or GPT‑6.1 Sol. The outer harness reviews an unfinished trajectory and decides whether to retry the task, end it or send it to a human.
In one example, Codex running GPT‑5.6 looked through thousands of recorded trajectories and spotted agents wasting tokens by repeatedly polling a backtest that took several hours. Codex then built and deployed a caching and memory-packing fix that cut the backtest from about 3 hours to 3–5 minutes.
GPT‑6.1 Sol is the lab's most-used model because it uses the fewest tokens while performing close to GPT‑6 Astra, according to a team member identified only as Adam. GPT‑6 Astra posted the biggest speedup on an internal latency-optimization benchmark. Cheaper models such as GPT‑5.6 Terra handle narrower jobs. The lab says it once used a trillion OpenAI tokens in a single day. It also uses OpenAI's Daybreak platform to security-test its agent environments before simulations launch.
This is a vendor case study. It gives no trading performance figures beyond a statement that the agents improved a baseline system.