Third Moment Research Deploys 10,000 AI Agents for 48-Hour Runs, Targeting Self-Improvement
Updated
Updated · OpenAI · Oct 14
Third Moment Research Deploys 10,000 AI Agents for 48-Hour Runs, Targeting Self-Improvement
2 articles · Updated · OpenAI · Oct 14
Summary
Third Moment Research is running 10,000 OpenAI-powered agents simultaneously on investigations lasting 36 to 48 hours as it builds a system that studies agent failures and automatically generates new training scenarios.
Codex, GPT-6 Astra, GPT-6 Sol and GPT-6.1 Sol power the setup, with an outer harness reviewing unfinished tasks to retry, stop or escalate them while subagents pursue parallel experiments.
One Codex-led optimization cut a quantitative-finance backtest on a year of U.S. equity order data from about 3 hours to 3 to 5 minutes, reducing token waste from repeated status checks.
Daybreak is used before simulations launch to test whether agents can escape their Dockerized research environments and reach the wider network, a safeguard Third Moment says is necessary at that scale.
The company says GPT-6.1 Sol is its most-used model for token efficiency, while its roadmap aims to expand from single investigations to hundreds of thousands of collaborating subagents.