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

Background and Context

OpenAI published a customer case study on Third Moment Research, a group that runs long-lived AI agents at scale. According to the public information, the organization operates about 10,000 agents at the same time. A single task can run for more than 36 hours. The stated goal is for agents to stay focused for days and to get better with use. The stack named in the case includes GPT-6 Astra, GPT-6.1 Sol, the coding agent Codex, and Daybreak, a product aimed at cybersecurity work. The case was published on October 14, 2026.

The source discloses only a few hard numbers: roughly 10,000 concurrent agents and runs beyond 36 hours. It does not describe the architecture, failure rates, or cost. Everything below that goes beyond those figures is analysis based on common industry practice. It is not a description of what Third Moment Research actually built.

Why does running an AI agent for 36 hours matter? A short task fits inside one context window. A task that runs for a day and a half does not. Context fills up, early decisions are forgotten, and tool errors pile up. One small deviation can grow into a directional error hundreds of steps later. That is the gap between a demo and a production workload.

Deep Analysis

How can an AI agent stay reliable for 36 hours or more? The source does not say how Third Moment Research does it. Systems of this type usually combine three capabilities. The first is externalized state. Plans, progress, and intermediate results are written to files or a database, not held only in conversation history. The second is checkpointing with resumable execution. After an interruption, the agent restarts from the last stable point instead of from zero. The third is periodic self-review. A second model instance or a separate process audits interim output and corrects drift early. The model lineup suggests a division of labor. GPT-6 Astra and GPT-6.1 Sol probably serve different reasoning or execution roles, though the source does not explain the split. Multi-model routing is a common way to control the cost of long tasks. A stronger, more expensive model handles planning and key judgments. A lighter model handles the many repetitive steps. Codex points to code as the working medium. Writing scripts, running tests, and reading logs all produce verifiable feedback. That feedback is what makes long autonomous runs practical. A test either passes or fails, and the agent can use that signal without a human in the loop. The phrase improve with use is also worth reading carefully. In engineering terms it most likely means saving successful and failed trajectories as reusable assets: prompt templates, tool wrappers, and evaluation cases. It does not necessarily mean retraining the model. The source does not confirm which method Third Moment Research uses.

Scale adds a second problem: scheduling and cost. Each of the 10,000 agents consumes inference quota and tool resources. A dead loop or pointless retry multiplies the bill. Budget caps, timeout reclamation, and circuit breakers become basic requirements. Rate limits, result deduplication, and task dependencies also need central management.

Industry Impact

The case moves the competitive question from the quality of a single answer to the reliability of sustained work. Vendors have spent the past year promoting coding agents that work for hours. A case measured in days, with concurrency in the thousands, invites enterprise buyers to use different metrics: task completion rate, cost per successful run, and recovery after failure. None of these is reported in the source, which is itself a limit on how far the case can be used as evidence. For OpenAI, the case supports a combined offer of models, Codex, and a vertical security product. It suggests the platform can support full long-running workflows, not just sell API access. Competitors such as Anthropic and Google, along with open-source agent frameworks and orchestration platforms, face pressure to show credible deployments at similar scale, not only demonstrations. The inclusion of Daybreak points toward cybersecurity work such as vulnerability review, code auditing, and threat analysis. These jobs split naturally into parallel pieces. A large codebase or asset list can be divided into many small sections, with one agent per section and a final aggregation step. This work demands accuracy and traceability, so log retention, permission isolation, and human review are usually essential. Longer runs also widen the attack surface for prompt injection and unauthorized tool use. Sandboxing and least-privilege design may matter more than raw model strength.

There is a dual-use side. Defenders can use agents for large-scale checks. Attackers could use similar capability to move faster. Access control and usage review will therefore become issues for platform providers.

Outlook

Several signals deserve attention. The first is whether the 36-hour mark is quickly surpassed, and whether vendors begin to publish shared benchmarks for long-horizon tasks instead of choosing their own showcase cases. The second is cost disclosure. The average spend per long task and the share of failed runs that must be repeated would show whether this model can move from a research organization to ordinary enterprises. The source gives neither figure.

The third is the mechanism behind getting better with use. It could rely on memory and skill libraries, on evaluation-driven prompt iteration, or on targeted fine-tuning. Each route carries different portability and different risks. The fourth is governance. Agents that run autonomously for days need audit trails and defined points for human intervention, especially in a sensitive field like cybersecurity. If these questions get public answers, long-running multi-agent systems could move from a special practice at a few organizations to repeatable infrastructure. The Third Moment Research case shows the direction. Maturity will still require more transparent data and stricter engineering validation.