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

Third Moment Research builds self-improving agents using GPT-6 Astra, GPT-6 Sol, GPT-6.1 Sol, Codex and Daybreak. It reports running 10,000 agents concurrently on investigations lasting 36–48 hours, consuming a trillion OpenAI tokens in a single day, and using Codex to cut a trading-strategy backtest from about three hours to three to five minutes — a 97% reduction.

An outer harness powered by GPT-6 Sol or GPT-6.1 Sol watches for unfinished tasks and decides whether to retry, terminate or escalate. Daybreak automates penetration testing of the research environments so agents can't break out into the wider network.

This is what multi-day agent research looks like when someone actually tries it: enormous scale, careful harnesses, and security testing of the sandbox itself. The backtest optimization story — Codex finding a caching bottleneck from recorded trajectories — is a concrete win, not a vibe.

We like that they treat unfinished work as a first-class failure mode with an outer harness that can escalate to humans. Long-running agents that just stop are worse than useless; they leave a false sense of progress.

The security piece is non-negotiable at this scale. Ten thousand agents with tools and data need breakout testing, and using Daybreak for that is the right instinct. Still, this is an OpenAI customer story with vendor-friendly framing; the finance results are company-reported.

Where this leads if it scales: companies that can afford trillion-token days pull further ahead on experimentation, and the gap between well-harnessed agents and casual chatbots widens.