OpenAI, in collaboration with the startup Third Moment Research, has triumphantly concluded a groundbreaking field test featuring a system of 10,000 intelligent agents. Throughout this rigorous evaluation, the agents functioned continuously for 36 to 48 hours, collectively processing up to a staggering 1 trillion tokens each day. The workload accomplished in a single round of tasks was equivalent to what a human engineer could achieve in 41 years. This test was conducted within a quantitative finance framework, where the agents autonomously optimized the underlying backtesting engine. Their efforts resulted in a remarkable 97% reduction in processing time and revealed anomalous patterns in dark pool trading activities at exchanges, which had previously eluded detection by human analysts.

To tackle challenges such as agents deviating from their intended paths over extended periods, escalating costs, and potential boundary transgressions, the system incorporates a sophisticated dual-layer supervision framework, a four-step self-iterative refinement process, and a robust sandbox security mechanism. The research team highlighted that quantitative investment and cybersecurity represent two domains with insatiable demands for computational resources. Presently, the system is progressively being deployed across a wider range of scenarios, with ambitions to elevate the concurrent agent capacity to the 100,000 mark. Simultaneously, the ability to effectively manage and enable self-iteration within large-scale agent clusters is emerging as a pivotal battleground in the competitive landscape of advanced AI models.