Third Moment Research runs 10,000 agents using GPT-6 models for investigations

Article Summary

Third Moment Research utilizes Codex and models including GPT-6 Astra, GPT-6 Sol, and GPT-6.1 Sol to operate 10,000 agents in parallel for investigations spanning 36 to 48 hours. By deploying Codex to analyze simulation trajectories, the company identified bottlenecks in quantitative finance backtesting, successfully reducing execution times from approximately 3 hours to 3 to 5 minutes through caching and memory optimizations. These models further support the research environment by enabling agents to spawn subagents and maintain coherent computational threads, while an outer harness manages task retries or escalations to human review when investigations remain incomplete.

AI Analysis

## Core Assessment
Third Moment Research is scaling agentic workflows by running 10,000 parallel agents across 36-to-48-hour windows, using GPT-6 variants and a Codex-based harness to manage task delegation and sub-agent spawning. The technical pivot here isn't just model scale, but the use of automated feedback loops: the system profiles agent trajectories to identify bottlenecks—such as the redundant polling that inflated token usage in financial backtesting—and then applies automated patches to fix them. Reducing backtest latency from three hours to five minutes is a legitimate engineering gain, as it clears context windows and minimizes "token waste." However, the system relies on an outer harness for error handling and human escalation, confirming that even with "self-improving" capabilities, the boundary between automated research and human oversight remains porous.

## Impact and Advice
* **Quantitative Researchers:** Leverage agentic profiling to audit your own research codebases. If your agents are spending more compute cycles on status checks or I/O bottlenecks than on model training, use the observed "caching and memory-packing" strategy to reclaim your token budget.
* **Systems Architects:** Before scaling to thousands of agents, implement an "outer harness" that handles task retries and state management. Without a dedicated supervisor to determine when to terminate a stalled thread or escalate to human review, your long-running investigations will likely drift into infinite loops of inefficient token consumption.

## Bottom Line
Automating the feedback loop between agent performance and code optimization is a sound strategy, but don't mistake 10,000 agents for 10,000 independent thinkers; they are only as effective as the harness managing their context.

Explore this story further