Inside Third Moment Research: 10,000 Agents Running for 36+ Hours!

Hello, it's Shii-chan! Today I found a large-scale story about running 10,000 agents at once!

OpenAI News openai.com

What was announced?

OpenAI News published a customer story about Third Moment Research, an AI lab. The company is building a self-improving AI system aimed at agents that are very good at a company's specific work and get better the more they are used.

Here is the loop.

• Understand agent behavior
• Identify recurring weaknesses
• Automatically create thousands of synthetic scenarios around those weaknesses
• Use the scenarios to improve agents, and repeat as new issues emerge

Codex, GPT-6 Astra, GPT-6 Sol, and GPT-6.1 Sol support the research and computational work. This lets the team run 10,000 agents at once on investigations lasting 36 to 48 hours.

Why it matters

Running many agents for a long time brings problems like wasted tokens, polluted context, and agents that stop early. This story is interesting because it shows concretely how the team handles them.

What changes

The clearest numbers in the story are about a faster backtest.

• Backtesting a strategy across a year of order-level US equity data took several hours
• While waiting, agents repeatedly polled for completion, wasting tokens and filling their context with status checks
• Codex found a caching optimization and a memory-packing technique and deployed the fix, cutting the run from about 3 hours to about 3 to 5 minutes

The saved token budget could go to further experiments.

Dive Deep

Here is how the setup works.

• The default Codex harness runs inside a Dockerized environment for quantitative finance research
• Agents receive a terabyte of data, documentation, and a repository with modules for feature engineering, GPU-supported model training, trading strategies, and simulation
• GPT-6 Astra, Sol, and 6.1 Sol provide the reasoning, while the Codex harness gives agents tools to plan, set objectives, spawn subagents, and compact context
• For agents that stop before finishing, Codex runs as a subagent inside an outer harness powered by GPT-6 Sol or 6.1 Sol. The outer harness examines the trajectory and decides whether to retry, terminate the job, or escalate to human review

The story also describes how models changed over generations. According to co-founder Adam, before GPT-5.5 models could only execute experiments prescribed by a user. With GPT-5.5 and GPT-5.6 they could design follow-up experiments from results. With the GPT-6 models, agents can investigate root causes through hundreds of parallel experiments with subagents.

Model choice also varies.

• GPT-6.1 Sol is the most-used model, thanks to token efficiency with performance similar to GPT-6 Astra
• On a benchmark about reducing a live scoreboard service's update latency, GPT-6 Astra delivered the biggest speedup
• For targeted work such as checks for specific security weaknesses in the simulation, the team uses the more efficient GPT-5.6 Terra

For security, the team uses Daybreak before simulations launch, having agents penetration-test whether an agent could break out of its research environment and reach the wider network.

Wrap-up

• Third Moment Research is building self-improving agents that get better with use
• Codex and GPT-6 models run 10,000 agents on 36 to 48 hour investigations
• A backtest dropped from about 3 hours to about 3 to 5 minutes through Codex optimizations
• An outer harness decides whether to retry, terminate, or escalate stalled agents
• Daybreak automates security testing of the environments at scale

A good read for anyone designing operations for long-running agents!