Third Moment Research Runs 10,000 Agents for 36-Plus Hours

Third Moment Research runs 10,000 agents for more than 36 hours at a stretch, according to the laboratory's own account of its work. The figure is the core claim of a field report the company published on its long-duration agent infrastructure, and it places the lab among a small group of organizations testing whether large language model agents can hold a task for days rather than minutes.

The scale claim carries two distinct numbers, and both matter. The first is the agent count: 10,000 agents running in a single sustained deployment. The second is duration: more than 36 hours of continuous operation. Most commercial agent deployments today run for minutes or a few hours before requiring human intervention, a restart, or a fresh context window. A deployment measured in days tests a different engineering problem altogether — memory management, error recovery and cost control at a scale that short-horizon agents never encounter.

Which models does the lab use?

Third Moment Research names four specific tools in its report:

• GPT-6 Astra
• GPT-6.1 Sol
• Codex
• Daybreak, which the lab deploys for cybersecurity work

The report describes the agents built on this stack with two defining properties: they "focus for days and improve with use." Both properties are worth parsing, because each corresponds to an open research problem.

Sustained focus — holding a task across dozens of hours without losing coherence — is the durability problem. Improvement with use is the adaptation problem: agents that get better at their assigned work as they accumulate experience on it, rather than returning to baseline performance at the start of every session.

Why does long-duration autonomy matter?

The stakes are economic as much as technical. Agents that persist for days can take on work that currently requires human shifts: extended security monitoring, sustained code refactoring, prolonged investigation of complex systems. The inclusion of Daybreak for cybersecurity indicates one concrete application domain Third Moment Research has in mind — monitoring and defense tasks where continuous coverage, not bursts of intelligence, is the binding constraint.

The choice of models also signals where the lab believes capability currently sits. GPT-6 Astra and GPT-6.1 Sol are the reasoning engines in the stack; Codex is the code-facing tool; Daybreak covers the security domain. A four-tool composition, rather than a single model, suggests the lab treats long-horizon agency as an orchestration problem, not a model problem — the durable system is the layer that coordinates specialized tools across a 36-hour run.

What is still unverified?

The 10,000-agent and 36-hour figures come from Third Moment Research itself. The report does not, in the material available, publish third-party benchmarks, task completion rates, cost per run, or failure statistics for the deployment. Independent verification of long-duration agent claims remains rare across the field, because few standard benchmarks exist for work measured in days. Readers should treat the numbers as the lab's own reported capability, not as an audited result.

What is concrete is the composition of the system and the stated design goal. Third Moment Research has named its models, named its cybersecurity tool, and stated its endurance target. If the lab publishes task-level results from these 10,000-agent runs — completion rates, cost curves, degradation measurements over the 36-hour window — the industry will gain one of the first usable datasets on how agent fleets actually behave at multi-day scale.