Third Moment Research runs 10,000 OpenAI agents for 36 hours

By Splitfeed EditorialLast updated Oct 14, 2026, 16:50 UTC

On October 14, 2026, Third Moment Research published a case study describing how it runs 10,000 simultaneous AI agents for sessions lasting more than 36 hours. The deployment draws on four OpenAI models: GPT-6 Astra, GPT-6.1 Sol, Codex, and Daybreak, with Daybreak specifically oriented toward cybersecurity tasks. The case study also claims that agent performance improves over the course of these extended runs, suggesting the system is not simply maintaining a steady state but developing in some measurable way during operation.

The scale described here is notable on its own terms. Running 10,000 concurrent agents for over a day and a half represents a significant operational commitment, and the use of four distinct specialized models within a single workflow points toward a more modular, multi-model architecture than earlier agentic deployments typically required. The inclusion of a cybersecurity-focused tool, Daybreak, raises questions that the case study does not fully answer, including what kinds of cybersecurity tasks are being automated, at what level of autonomy, and with what oversight mechanisms in place. The claim that agents improve with use over extended sessions is also a point that invites scrutiny, since the nature and verification of that improvement are not detailed in the available material.

No published reactions from the Pro-AI, Anti-AI, or Middle Ground camps have appeared yet. Pro-AI voices would typically treat a deployment at this scale as evidence that agentic AI is maturing into a reliable research and operations tool, pointing to the sustained runtime and multi-model coordination as proof of concept. Anti-AI commentators would likely focus on the cybersecurity angle, arguing that autonomous agents operating at this scale in sensitive domains warrant much closer examination of risks and accountability structures. Middle Ground observers would probably call for transparency about what the agents are actually doing, how improvement is being measured, and what safeguards govern a system of this size.

The details most worth watching are any technical disclosures Third Moment Research or OpenAI release about how agent improvement is defined and measured across these sessions, and whether further information emerges about the specific scope and boundaries of Daybreak's role in the workflow.