10,000 AI Agents Run for Days: Third Moment's OpenAI Breakthrough

Autonomous Research at Unprecedented Scale

The article showcases an impressive feat of running 10,000 AI agents concurrently for extended durations, leveraging advanced OpenAI models like GPT-6 Astra and GPT-6.1 Sol, alongside Codex for tooling and Daybreak for security. The core innovation lies in Third Moment's self-improving AI system, which uses these models to analyze agent behavior, identify weaknesses, and synthetically generate training data. This iterative process allows agents to become more capable over time, a critical step towards more robust and autonomous AI systems. The reduction in backtesting time from hours to minutes using Codex is a tangible demonstration of efficiency gains, directly impacting development cycles and token waste. This approach has clear benefits for quantitative finance, but the article also hints at broader applications in supply chain optimization and code analysis, suggesting a versatile framework.

However, the article is light on the specifics of the 'self-improvement' loop beyond identifying weaknesses and creating synthetic scenarios. The exact mechanisms for how agents learn and adapt from these scenarios are not deeply explored. Furthermore, while GPT-6.1 Sol is praised for token efficiency, the article doesn't quantify the cost implications of running such a massive number of agents, even with efficiency gains. The reliance on proprietary OpenAI models also introduces vendor lock-in, a potential concern for long-term scalability and cost management. The article implies a significant computational infrastructure is required, but details are scarce. Despite these points, the demonstration of sustained coherence and task completion across thousands of agents over days represents a significant leap forward in AI agent orchestration and autonomous research.

Key Points

• Third Moment Research runs 10,000 AI agents concurrently for 36-48 hours using OpenAI's GPT-6 models and Codex.
• The company employs a self-improving AI system that analyzes agent behavior, identifies weaknesses, and generates synthetic scenarios for iterative improvement.
• Codex significantly reduces backtesting time (e.g., from 3 hours to 3-5 minutes) and token waste by optimizing research environments.
• GPT-6 models (Astra, Sol, Sol) provide reasoning, while Codex offers tooling for agents to act, plan, and spawn subagents.
• Daybreak is used for automated security testing of the research environments.
• The system enables agents to explore research directions autonomously, uncovering unexpected insights (e.g., impact of non-displayed trades).
• Future plans include scaling to hundreds of thousands of collaborating agents for complex, multi-stage research.

---

📖 Source: How Third Moment Research runs 10,000 agents for 36+ hours