Ten thousand agents running for more than 36 hours is a duration claim, not a headcount claim, and that is what makes the OpenAI customer story about Third Moment Research worth reading. The post, dated October 14, 2026, names GPT-6 Astra and GPT-6.1 Sol alongside Codex and Daybreak for cybersecurity, and says the lab builds agents that focus for days and improve with use.
Key facts
| Fact | Detail |
|---|---|
| Question answered | Third Moment Research Runs 10,000 Agents for 36+ Hours on OpenAI Models |
| Short answer | OpenAI published a customer story on October 14, 2026 describing how Third Moment Research runs 10,000 agents for more than 36 hours at a time, using GPT-6 Astra and GPT-6.1 Sol together with Codex and Daybrea… |
| Reading time | ~2 min |
| Last updated | 2026-10-15 |
What is actually on the record
The verifiable part is short. The publisher is OpenAI, the date is October 14, 2026, the customer is Third Moment Research, and the summary line states the 10,000-agent, 36-hour-plus figure and the model mix. No cost per run, success rate or failure taxonomy is published in the text we could read, and the page returns HTTP 403 to this machine.
Duration is the hard part
A single-session agent can be sloppy about state: if it crashes, you rerun it. An agent fleet that is supposed to stay useful for a day and a half cannot work that way:
- it needs checkpoints it can resume from, because nothing stays up for 36 hours by accident;
- it needs a spend ceiling that survives a restart, or an overnight loop becomes an unbounded invoice;
- it needs logs detailed enough that a human can reconstruct what happened while asleep;
- it needs a way to stop, which means a policy for irreversible actions rather than a keyword filter.
Those four are engineering choices, and they are the same choices whether you run ten agents or ten thousand.
“Improve with use” is the other claim
The second half of the summary is that the agents improve with use. In practice that means a harness that feeds outcomes back into the next run - rules, skills, or prompts updated from what failed last time rather than from a developer’s guess. The portable version of that idea costs nothing: keep the failures, and write down the rule that would have prevented each one.
A short checklist
- Set the budget before the run, not after the first surprise.
- Verify that your agent can resume from a checkpoint instead of restarting from zero.
- Keep an approval gate on anything irreversible.
- Log per-step model and token usage so a long run can be priced afterwards.
For agents on your own Windows machine, where the files never leave the disk unless you say so, download MOOGH and start with one overnight task rather than ten thousand.
Related reading: