Third Moment Research is running 10,000 autonomous agents at once in sessions that last 36 to 48 hours (OpenAI). The multi‑day workloads consume up to one trillion OpenAI tokens in a single day as the lab builds self‑improving agents for company‑specific environments.
The architecture leans on OpenAI models—GPT‑6 Astra, GPT‑6 Sol and GPT‑6.1 Sol—for core reasoning. Third Moment notes that GPT‑6.1 Sol can keep a coherent research thread across days of compute without losing context. In addition, Codex handles agent tooling while Daybreak tests the cybersecurity of operational environments.
To sharpen agent behavior, the lab examines execution trajectories, isolates recurring failure modes and creates thousands of targeted training scenarios around those gaps. It compares recorded agent actions and timestamps against the codebase to flag structural inefficiencies before retraining and redeploying agents on updated workflows.
One early deployment focuses on quantitative finance, chosen for its clear benchmarks. While probing early bottlenecks with Codex and GPT‑5.6, the team found that backtests on annual order‑level equity data ran about three hours. Continuous polling during that window exhausted token limits and flooded context windows with repetitive status queries.
Codex addressed the problem by profiling inputs and adding memory‑packing techniques plus caching optimizations directly into production. The fix cut backtest latency by 97 percent, shrinking runtimes from roughly three hours to three to five minutes. This curbed context pollution and freed token budgets for live experiments.
For quantitative tasks, the team runs the default Codex harness inside Dockerized environments. Each agent accesses a terabyte of reference data, system documentation and a codebase that includes modules for feature engineering, GPU‑accelerated model training and market simulation.
The integrated Codex harness equips autonomous agents with tooling to plan workflows, set milestones, trigger subagents and dynamically compact working context throughout multi‑day runs.
The massive scale of one trillion tokens daily across 10,000 agents suggests that operational bottlenecks in autonomous computing are shifting from reasoning limits to infrastructure latency. The dramatic drop in backtest times indicates that context pollution caused by system polling remains a major point of failure for persistent agents. Consequently, long-duration autonomous workflows appear increasingly reliant on aggressive context compaction and localized tool optimization rather than expanded raw model capacity alone.
Spotted an error? See our corrections policy or email editorial@lordofclaude.com.
Discussion & Comments (0)
Join the conversation