Large-Scale Autonomous Investigations
On October 14, 2026, Third Moment Research unveiled a system capable of executing extensive, multi-day autonomous investigations using various OpenAI models. The company reports that it can run 10,000 agents simultaneously in sessions lasting 36 to 48 hours, processing significant volumes of tokens in production.
Agent Scalability and Task Management
Agents operate within a Dockerized environment, equipped with a terabyte of data, documentation, and a research repository that includes modules for feature engineering, GPU-supported model training, trading strategies, and simulations. Third Moment employs Codex as an action-oriented harness, enabling agents to plan, set objectives, spawn subagents, and compact context. Meanwhile, GPT-6 Astra, GPT-6 Sol, and GPT-6.1 Sol provide higher-level reasoning capabilities.
To mitigate the risk of unfinished tasks, the company implements an outer harness driven by GPT-6 Sol or GPT-6.1 Sol. This outer layer inspects an agent’s trajectory and determines whether to retry, terminate, or escalate the task to human review. Adam Khakhar, co-founder of Third Moment Research, stated, “The Codex harness is extremely versatile, and it enables use across many different domains.”
Performance Enhancements and Security Testing
One significant outcome arose from a quantitative finance workload. Third Moment utilized Codex to analyze thousands of simulation traces, identifying a bottleneck in backtesting. By applying a caching optimization and a memory-packing technique identified by Codex, the company successfully reduced the backtesting duration from approximately three hours to around three to five minutes. This improvement notably decreased token waste and context pollution.
Jeffrey Yan, co-founder, mentioned, “GPT-6.1 Sol is currently our most-used model given its token efficiency. It completes tasks with the fewest tokens while delivering performance similar to GPT-6 Astra.” The team is also testing with GPT-5.6, including GPT-5.6 Terra, and evaluating GPT-6 Luna for various workloads.
For automated security checks at scale, Third Moment employs the Daybreak platform, allowing agents to perform penetration tests on simulation environments and verify isolation before launch. Adam Khakhar emphasized, “Without Daybreak, we would not be able to autonomously test this at scale.”
Future Plans for Collaboration and Experimentation
Third Moment’s roadmap aims to transition from individual investigations to hundreds of thousands of collaborating subagents. The goal is to support chains of dependent experiments backed by trillions of tokens, utilizing these experiments to autonomously enhance agents as new challenges arise.
Original source: OpenAI News