How Third Moment Research Runs 10,000 AI Agents

Background and Context

OpenAI recently published a customer story about Third Moment Research, an AI lab that runs agents at very large scale. The lab uses GPT-6 Astra, GPT-6.1 Sol and Codex. It also uses Daybreak, a product aimed at cybersecurity work. The story gives two headline numbers: about 10,000 agents, and runs that last more than 36 hours. The stated goal is an agent that stays focused for days and gets better with use.

The story appeared in mid-October 2026. Vendors are now moving agents from demos into production, and this case fits that shift. The aim is not one impressive demonstration. The aim is for agents to work like long-term staff, who keep what they learn and do better on the next round.

One limit matters here. The public summary does not name the task types, the success rates or the cost. This article therefore reads the story through what is published and through well-known engineering practice. It does not guess at figures that the source leaves out.

Deep Analysis

How does Third Moment Research keep 10,000 AI agents running for more than 36 hours? The source does not describe the method in detail. It names the tools: GPT-6 Astra, GPT-6.1 Sol, Codex and Daybreak. It states the scale: about 10,000 agents and runs beyond 36 hours. Beyond that, any mechanism is inference from common practice, not a disclosed fact. Long runs and short chats are different engineering problems. A run of a few minutes tests the reasoning quality of the model. A run of 36 hours or more tests durability. The context window fills with tool output and draft work. Early goals fade behind late details. Agents can loop on the same failed attempt, or drift quietly from the original task. Teams usually answer these risks with several measures together: split the job into stages that can be checked, write key state to external memory, save checkpoints so a run can resume after a break, and add an independent step that checks the work.

The named products suggest a division of labour. Codex works in code environments. It reads and writes files, runs commands and iterates, so it fits the execution layer. GPT-6 Astra and GPT-6.1 Sol are two different models. They may take planning and high-volume execution roles, although the source does not say so. Sending easy steps to a lighter model is a common way to hold cost down at this scale. The phrase improve with use also needs care. It does not have to mean new model weights. More often it means that successes, failures and reusable tool routines are saved in a knowledge base, prompt templates or a skill library. Later agents then inherit them. Here a large fleet helps: with 10,000 instances, the system can see sooner which strategies hold up. Daybreak is described only as a cybersecurity tool. Security work suits long-running agents, because bug hunting, log review and attack-surface mapping need steady exploration and repeated checks. The same work also demands tight permissions and a clear audit trail.

Industry Impact

For model vendors, the contest is moving away from single benchmark scores. Run length, parallel scale and stability now count for more. A vendor that offers dependable long runs is better placed to become the base platform for enterprise agents. OpenAI chose a customer story over a parameter table to make this point. That choice signals that it wants to prove the whole stack of models plus tools under a real workload, not one model alone.

For other labs and start-ups, the bar is rising. Building an agent used to centre on prompts and tool wiring. Now the team must also handle scheduling, fault tolerance, cost accounting and observability. This looks more like distributed systems work than prompt writing. Rival model makers and open-source agent frameworks are pushing toward long tasks too. They will differ on running cost, depth of fit with existing developer tools, and how well they handle failure.

For cybersecurity, a purpose-built agent like Daybreak can automate a large share of repetitive investigation for defenders. The same capability could help attackers. Access control, usage audit and clear lines of responsibility will therefore shape how such products are built. For ordinary enterprise buyers, the lesson is practical. Do not ask only how smart the model is. Ask whether the system can finish a task that spans days, whether it can roll back after an error, and whether every step can be traced.

Outlook

Several points deserve attention next. First, will Third Moment Research publish firmer metrics? Useful ones include task completion rate, how often a human steps in, and the average cost per agent. It would also help to learn whether runs can stretch beyond 36 hours to several days. Without these numbers, no one can tell if long-running agents are a repeatable engineering skill or a result that holds only in narrow cases. Second, how will the improve-with-use loop work in practice? If it relies on external memory and saved skills, the key risk is that a bad lesson gets fixed in place and spreads across the whole fleet. If it involves retraining the model, data governance and evaluation methods become the main questions. Third, scale brings safety and compliance duties. These include least-privilege permissions for each agent, complete action logs and a circuit breaker for abnormal behaviour. They matter most in a high-risk field such as cybersecurity. Fourth, cost will shape the design. Running 10,000 agents for 36 hours or more uses a great deal of compute. Mixing model sizes and handing fixed steps to plain scripts may be the main ways to cut the bill.

If these questions get good answers, long-running agents could move from the lab into daily enterprise work in software development, security operations and research. If reliability and cost stay out of reach, buyers will likely fall back to more cautious human-and-agent teamwork. Either way, the case shows that the next stage of the agent contest will be decided by durable, coordinated and auditable engineering.