Third Moment Research Scales 10,000 Autonomous Agents for Multi-Day Operations

Third Moment Research has disclosed operational details of a large-scale agent deployment that kept 10,000 autonomous agents active for more than 36 continuous hours. The workload leverages GPT-6 Astra, GPT-6.1 Sol, Codex, and the Daybreak cybersecurity suite to perform extended reasoning, code generation, and threat-hunting tasks that improve iteratively with each cycle.

Why it matters for GPU / AI infrastructure

Sustained, multi-day agent fleets shift the cost profile from burst inference to persistent compute. Operators need GPU clusters that deliver high memory bandwidth, low tail latency, and predictable pricing over weeks — not hours. This pattern favors reserved-instance pools, disaggregated memory architectures, and scheduling layers that can preempt lower-priority batch work without starving long-running agents.

The lab reports that agent performance compounds over time as each cycle refines context windows and tool-use policies. That feedback loop demands checkpointing and fast model reloads, putting pressure on NVLink-domain scale-up and high-throughput object stores for state serialization.

For cloud providers and enterprise buyers, the takeaway is clear: capacity planning must account for workloads that never fully idle. Pricing models, quota systems, and autoscalers designed for request-response traffic will under-provision for persistent agent swarms.