AI Lab Runs 10,000 Agents for 36+ Hours, Advancing Long-Running AI

An AI research lab has successfully orchestrated a swarm of 10,000 AI agents to run continuously for more than 36 hours, demonstrating a significant leap in the scale and endurance of autonomous AI systems. This experiment, conducted by Third Moment Research, points toward a future where large populations of AI can perform sustained, complex work without human intervention.

According to a post on OpenAI's news site, the lab utilized a suite of advanced models including GPT-6 Astra and the more efficient GPT-6.1 Sol, alongside Codex and a specialized cybersecurity model called Daybreak. The core achievement was not just the number of agents but their ability to maintain focus on long-horizon tasks, learning and improving their performance over the extended runtime. This work moves beyond simple, one-off AI interactions into the realm of persistent digital workforces.

How did they run so many agents for so long?

The technical feat hinges on combining efficient model inference with robust orchestration software. Third Moment Research used GPT-6.1 Sol, a model optimized for lower-cost, high-volume tasks, as the workhorse for many agents. This allowed them to manage the substantial computational load of 10,000 concurrent agents. The lab's own custom framework handled the lifecycle of these agents—spawning them, managing their state, facilitating communication, and ensuring they remained on task. The 36-hour duration, a significant increase over typical short-lived agent experiments, was made possible by this stable infrastructure and the models' inherent reliability.

What can these long-running agents actually do?

The agents were deployed on a variety of research-oriented tasks that benefit from prolonged, iterative effort. While specific applications were not detailed, the source indicates a focus on cybersecurity simulations using the Daybreak model, along with complex code generation and analysis tasks powered by Codex and the GPT-6 family. The key capability is that these agents don't just answer a prompt and stop; they can be set on a multi-step problem, work on it, learn from intermediate results, and adapt their approach—all over a period of days. This mimics a form of sustained research or development cycle.

Why does running agents for days matter?

Most AI agent demos today are ephemeral, solving a single problem in a single session. The ability to run thousands of agents for days unlocks new possibilities. It means AI can be deployed for longitudinal studies, continuous system monitoring, or open-ended creative and research projects where the path isn't fully known upfront. The lab noted the agents "improve with use," suggesting they refine their strategies over time, leading to better outcomes. This endurance is a prerequisite for AI that can manage real-world processes, which rarely conclude in a five-minute chat session.

How does this relate to cost and efficiency?

Scaling to 10,000 agents would be prohibitively expensive with less efficient models. The use of GPT-6.1 Sol, noted for its cost-effectiveness in agentic workloads, was crucial. This aligns with a broader industry trend of optimizing model selection for specific tasks to control expenses, as seen in a recent case where Asana cut AI browser agent costs 76x with GPT-6.1 Sol. Third Moment's work demonstrates that for large-scale, persistent agent systems, model efficiency is as important as raw capability.

How can you explore autonomous agents in Grafy?

While few creators will need to spin up 10,000 agents, the principles of persistent, task-oriented AI are accessible. Grafy's Operator app is a browsing agent designed to perform multi-step web tasks autonomously, embodying the concept of a single, focused AI agent. For those interested in the research behind agent scaling and efficiency, Grafy Reader is an ideal tool to upload and interrogate technical papers or reports on the subject. Furthermore, the concept of multi-agent collaboration can be explored in Grafy Chat using features like the Council of Models, where different AI models can debate a problem, simulating a simplified agent network.

What are the broader implications for AI development?

This research is a concrete step toward the "agentic" future often predicted for AI. It provides a proof-of-concept that large-scale, long-duration AI systems are technically feasible today. However, it also surfaces important questions about oversight, cost management, and the measurable output of such systems. A recent study found AI coding agents increase code volume but not software output, a reminder that scale alone does not equate to useful productivity. As these systems evolve, the focus will need to shift from pure scale to effective governance and outcome measurement.

Frequently asked questions

How is this different from just using an AI API many times?

Running 10,000 separate API calls is stateless. These agents maintain persistent identity, memory, and objectives over 36 hours. They can learn from prior actions within their session, collaborate, and pursue long-horizon goals that require many iterative steps, which a series of independent calls cannot achieve.

What hardware is required for 10,000 agents?

The source does not specify hardware details. The achievement likely relies heavily on cloud infrastructure and efficient model serving. The key is software orchestration and using cost-optimized models like GPT-6.1 Sol to make the massive parallel workload economically and technically viable.

Could this technology be used for creative projects?

Yes. While the lab focused on research and cybersecurity, the architecture could be adapted for creative fields. Imagine a swarm of agents generating variations on a theme, simulating audience reactions, or managing the long-term narrative arc of an interactive story over days, similar to planning a project in Grafy's Graph but at a much larger scale.

---

Source: OpenAI News — How Third Moment Research runs 10,000 agents for 36+ hours