10,000 Codex agents ran for 48 hours, rewrote their own backtest engine for 97% speedup
新智元 · wechat · 2026-10-15
- OpenAI showcased a stress test with Third Moment Research: 10,000 agents ran concurrently for 36-48 hours, consuming 1 trillion tokens daily — a single round equals 41 engineer-years.
- The agents found backtesting on a year of order data took 3 hours; Codex independently located two bottlenecks and shipped cache and memory-packing optimizations, cutting runtime to 3-5 minutes (97% faster).
- GPT-6.1 Sol chased an anomaly to a dark-pool liquidity pattern at an exchange that the human quant team had overlooked.
- A dual-layer architecture wraps Codex workers with a supervisor model (GPT-6 Sol / 6.1 Sol) that decides retry/terminate/escalate when inner agents stall; a four-step loop (trajectory collection, clustering via GPT-6 Luna, synthetic sandboxes, scaffold upgrades) continuously improves the system.
- Security is handled by OpenAI's Daybreak red/blue teaming agents plus GPT-5.6 Terra; the company argues quant and cybersecurity are the only zero-sum fields with unlimited compute appetite, and plans to scale to hundreds of thousands of agents.
More from coding & agent
- Claude Code Pushes Urgent Security Update Across All Installations — umeshonai · 2026-10-23
- Claude cut a 2.6-hour video export to 20 minutes — but only after being told to rebuild the server — byhewar · 2026-10-23
- 3 flight-booking rules my AI agent checks that I'd never do by hand — Efistoffeles · 2026-10-23
- OpenJet claims Qwen3.8-27B coding runs 1.9x faster than llama.cpp on one RTX 3090 — JeremyCMorgan · 2026-10-23
- Building an autonomous SEO agent in Hermes with three skills: programmatic pages, Keywords Everywhere, DataForSEO — EXM7777 · 2026-10-23
- Building a sports betting AI agent with onchain execution and human approval by default: architecture and lessons — Dry-Comfortable-2678 · 2026-10-23