What the team built
Third Moment Research describes an agent system that can run 10,000 agents in parallel for more than 36 hours. The company says a single day of activity processed one trillion tokens. That is a striking engineering stress test: the agents divide work across a long-running research operation instead of answering one short prompt.
The same case study gives a more grounded example. Codex reduced a backtest over a year of U.S. equity order-level data from about three hours to three to five minutes, according to Third Moment. That specific benchmark is roughly a 97% time reduction.
Why it matters—and what it does not prove
The valuable change is throughput: parallel agents can explore many research paths while a human team defines the task and judges the result. It suggests software for orchestrating agents may matter as much as raw model capability.
These are company-reported results, not an independent benchmark. The case study does not establish that 10,000 agents improve accuracy or that the backtest translates into trading returns. The clearest claim is that the workflow completed this particular data job much faster.
What to watch
A large swarm is useful only when tasks can be divided cleanly, progress observed and outputs reconciled. Third Moment’s example puts attention on that orchestration layer, not merely on one model’s speed. For researchers, the opportunity is running more experiments in the same time window. For finance readers, the backtest is not investment advice and says nothing about future returns. Related context: https://miketechlife.com/article/openai-eu-agentic-use-grows-8x
Source published 2026-10-14. Coverage is based on the maker’s announcement and demonstration.