Third Moment Research Runs 10,000 AI Agents for 36 Hours Using OpenAI Models

OpenAI has published first-hand research describing how the AI lab Third Moment Research successfully managed a large-scale deployment of autonomous agents. The report details the technical approach used to run 10,000 distinct agents for a duration exceeding 36 hours without manual intervention.

The study highlights the operational capabilities of specific OpenAI models in sustaining long-term autonomous workflows. Third Moment Research relied on GPT-6 Astra and GPT-6.1 Sol to maintain agent focus over multi-day periods. These models enabled the agents to improve their performance through continuous use during the extended testing window.

Technical Implementation and Model Usage

The research specifies that the lab integrated multiple model versions to handle different aspects of the autonomous workload. In addition to the primary GPT-6 variants, the team utilized Codex and Daybreak models. According to the OpenAI publication, these additional models were specifically directed toward cybersecurity tasks within the agent ecosystem.

The primary technical achievement documented is the ability of these agents to maintain focus for days. The agents demonstrated an capacity to improve with use, suggesting that the underlying models can adapt to prolonged operational contexts. This setup allows for complex, multi-step reasoning that persists beyond short-term interaction limits.

Scope of Agent Capabilities

The deployment involved a significant volume of concurrent processes, totaling 10,000 individual agents. Each agent operated autonomously for over 36 hours, indicating a high level of system stability and resource management by the research team. The focus of the agents was not limited to general tasks but included specialized cybersecurity operations.

The use of Daybreak and Codex for cybersecurity suggests a targeted application of these models for defensive or analytical security purposes. The research does not specify the exact nature of the cybersecurity tasks beyond their inclusion in the agent workflow. The emphasis remains on the duration and scale of the autonomous operation rather than specific security outcomes.

What should you do?

Security and technology professionals should monitor developments in autonomous AI agent capabilities, particularly regarding long-duration operations. Understanding how models like GPT-6 variants sustain focus can inform risk assessments for systems that may interact with or deploy similar agents. Organizations should evaluate their current infrastructure for potential exposure to automated, long-running AI processes.

What to do and how to stay safe: OpenAI

• Review current policies regarding the deployment of autonomous AI agents in production environments.
• Assess the potential impact of long-running AI processes on system resources and security monitoring.
• Monitor vendor updates regarding the capabilities and limitations of new model versions like GPT-6.
• Ensure that logging and audit trails are robust enough to track actions taken by autonomous agents.

Step-by-step guide: Prompt Injection Warning Signs: How to Spot AI Manipulation

General security guidance from the Threat Times newsroom. It is not confirmed advice from the organisations named in this story.

Frequently asked questions

How many agents did Third Moment Research run?

The lab ran 10,000 agents continuously for more than 36 hours, according to the OpenAI research.

Which AI models were used for the cybersecurity tasks?

Codex and Daybreak models were used for cybersecurity tasks, while GPT-6 Astra and GPT-6.1 Sol supported general operation.

Did the agents improve during the test?

Yes, the research states that the agents improved with use during the multi-day operational period.