Third Moment Research Runs 10,000 AI Agents for 36+ Hours

Background and Context

Third Moment Research is an AI lab. According to a case study published by OpenAI, it is building agent systems on OpenAI models and tools that can work for a long time without a break. The headline figure is large: more than 10,000 agents running at the same time, for over 36 hours. The stack named in the case has four parts. There are two models, GPT-6 Astra and GPT-6.1 Sol. There is Codex, the coding agent. And there is Daybreak, which targets cybersecurity work.

The lab states its goal plainly. It wants agents that stay focused for days and improve with use, instead of resetting to zero when a task ends. One caveat matters. The public material does not say how the work is split between agents, what tasks they ran, what the run cost, or how often they succeeded. Everything below that goes past these facts is analysis based on common engineering practice. It is not a description of the team's internal design.

Deep Analysis

The 36-hour figure is notable because most agent products still work in short bursts. A typical job lasts minutes to tens of minutes. The model plans, calls tools and returns a result inside one context window. Stretch that to dozens of hours and four problems appear. The first is context drift. Early goals and constraints get diluted by later information, and an agent can slide away from its brief without anyone noticing. The second is error accumulation. Each step carries some chance of failure, so reliability falls as the chain grows. Long runs need checkpoints, rollback and self-checks. The third is state. No run of this length fits in one conversation, so progress must live in external memory, task logs and file systems. An interrupted agent can then resume where it stopped. The fourth is scheduling. With 10,000 agents, there are huge numbers of concurrent model calls, tool executions and sandboxes. A rate limit or fault in any one layer can spread across the whole fleet.

Surviving 36 hours therefore suggests that the orchestration and fault-tolerance layer has reached a useful level of maturity. The phrase "improve with use" also needs care. In practice it usually means that lessons from earlier runs are saved as reusable memory, skills or evaluation data, so later agents waste less effort. That is learning at the system level. It does not necessarily mean that model weights change, and the source does not say which one Third Moment means. The pairing of components hints at a division of labor. The large models handle reasoning and decisions. Codex turns decisions into programs, scripts and automated workflows. Daybreak points the capability at security, a field with high value and high risk.

Industry Impact

Cybersecurity is a natural test bed for long-running agents. Vulnerability hunting, log analysis and attack-and-defense exercises all reward patience, repetition and wide coverage. Human time is expensive in this work, while machines can run without rest. The same setting also demands tight control. Every action an agent takes should be auditable and limited in scope. Otherwise more capability simply means more risk.

For developers and enterprise buyers, the yardstick may shift. Today the focus is the quality of a single answer. A case like this moves attention to stability over a long run, ability to recover after a failure, and cost per completed task. For rival model and agent vendors, long-horizon work becomes a new place to compete. Benchmark scores show little about it. What persuades buyers is whether a system can keep a complex job going until it is done. For the security industry, the effect cuts both ways. Defenders can use fleets of agents for large-scale inspection. Attackers could use similar capability to work faster. That raises the importance of access control, usage review and responsible release for any security-focused tool, Daybreak included.

Outlook

Four signals deserve watching. The first is whether OpenAI turns this kind of large-scale, long-duration orchestration into a product. Task queues, memory stores, checkpoints and monitoring interfaces would spare other teams from building them from scratch.

The second is cost. Running 10,000 agents for 36 hours implies a heavy compute bill. The model can only leave the lab if the cost per task falls to a level ordinary companies accept. No figures have been published yet. The third is evaluation. The field needs benchmarks that measure completion of tasks lasting days, and a trustworthy way to tell whether an agent is really getting better or only appears to. The fourth is governance. Authorization, audit trails and abuse prevention for security tools such as Daybreak will say a lot about how responsibly this capability spreads. The direction of the case is clear. Competition among agents is moving from who can answer well to who can finish the job, reliably and safely, over a long time. Until more details are released, one case should not be read as the industry norm.