Third Moment Runs 10,000 AI Agents for Days

Third Moment Research says it ran 10,000 AI agents continuously for more than 36 hours, using a mix of GPT-6 Astra, GPT-6.1 Sol, Codex, and Daybreak for cybersecurity. The lab describes the experiment as a demonstration of long-horizon autonomy: agents that can stay focused for days, operate in parallel, and improve as they work. That is a notable step beyond short, single-task AI interactions. If such systems can be sustained reliably, they could change how research teams, security operations centers, and software groups handle large workloads. Thousands of agents could monitor networks, test code, review documents, or explore hypotheses at once, with human supervisors setting goals and checking results. However, the experiment also raises hard questions. Running thousands of agents for long periods requires significant compute, energy, and coordination. More importantly, oversight becomes difficult when swarms act simultaneously and pursue subtasks that humans may not review in real time. Reliability is another concern: an agent that improves with use may also repeat mistakes, drift from instructions, or interact with other agents in unpredictable ways. Security is especially sensitive because the setup includes cybersecurity tools and models. If autonomous agents can operate for days, organizations will need clear rules for permissions, logging, escalation, and shutdown. Third Moment's claim may not settle those debates, but it points toward a future in which persistent AI workforces are tested in real settings. The value will depend on whether the approach delivers measurable gains without creating new risks that outweigh the benefits.