How Third Moment Research runs 10,000 agents for 36+ hours

Third Moment Research builds self-improving AI agents tailored to specific workflows, using OpenAI’s latest models to analyze behavior, identify weaknesses, and generate synthetic scenarios for continuous improvement. These agents operate autonomously for 36 to 48 hours, supported by Codex, GPT‑6 Astra, GPT‑6 Sol, and GPT‑6.1 Sol, while Daybreak tests the security of their research environments. The approach emphasizes domain-specific optimization, with early applications in quantitative finance where success is measurable.

In quantitative finance research, Third Moment’s agents analyze simulation trajectories and codebases to uncover inefficiencies. Using Codex with GPT‑5.6, the team identified a bottleneck where backtesting a trading strategy took hours, wasting tokens and polluting context. A caching optimization and memory-packing technique reduced backtest time from roughly 3 hours to 3–5 minutes, cutting token use and enabling more experiments.

Agents operate within a Dockerized environment containing a terabyte of data, documentation, and tools for feature engineering, model training, and strategy simulation. GPT‑6 Astra, GPT‑6 Sol, and GPT‑6.1 Sol provide reasoning, while the Codex harness enables planning, tool use, and context management. An outer harness powered by GPT‑6 Sol or GPT‑6.1 Sol handles incomplete tasks by retrying, terminating, or escalating to human review.

Third Moment evaluates models across optimization, machine learning, and quantitative research, testing frontier releases at high reasoning settings. GPT‑6 Astra delivered the largest speedup in reducing latency for a live scoreboard service while maintaining correctness. The team also uses GPT‑5.6 Terra for targeted security checks and plans to expand from individual investigations to collaborative subagents running trillions of tokens.