Third Moment AI agents scale to 10,000 concurrent runs for 36-hour sessions
Third Moment AI agents now launch 10,000 autonomous agents simultaneously for 36-hour research cycles, a jump from earlier pilots that capped at a few hundred agents. The platform relies on OpenAI’s GPT-6 Astra, GPT-6 Sol, and GPT-6.1 Sol for reasoning, and Codex for tool-based execution.
Codex slashes backtest duration from hours to minutes
In quantitative-finance experiments, agents backtested trading strategies against a year of order-level US equity data. Codex, paired with GPT-5.6, analyzed thousands of simulation trajectories and identified a caching bottleneck that forced agents to poll for completion, consuming tokens and context space. Codex suggested a memory-packing fix and a caching layer, which the team deployed. The backtest time dropped from roughly three hours to three-to-five minutes, cutting token waste and freeing context for further experimentation.
GPT-6 maintains coherent research threads over days
A co-founder of Third Moment noted that GPT-6.1 Sol can keep a coherent research thread across days of compute without losing the plot. The model processes long-term dependencies in the agents’ logs, allowing agents to revisit earlier decisions without re-deriving context. This capability is critical when agents run for 36-48 hours, as it prevents drift and reduces the need for manual intervention.
Agent environment and tooling
Each agent operates inside a Dockerized harness that bundles a terabyte of data, documentation, and a research repository. The repository includes modules for feature engineering, GPU-accelerated model training, trading strategies, and simulation. Codex supplies the agents with a toolset that can read, write, and execute code, while GPT-6 models provide high-level reasoning.
Security testing with Daybreak
Third Moment also employs Daybreak to audit the security of the environments where agents run. Daybreak scans for vulnerabilities in the Docker images, network configurations, and data pipelines, ensuring that the agents operate in a hardened sandbox.
Scaling roadmap
The company signals a move toward hundreds of thousands of collaborating agents. Achieving that scale will require further reductions in token waste, tighter integration of caching mechanisms, and more efficient context management. The current 97% reduction in backtest time demonstrates a path forward, but sustaining longer research sessions will demand additional optimizations.
Implications for developers
Developers building autonomous systems can learn from Third Moment’s approach: combine a large-language model for reasoning with a code-generation model for tool execution, and use profiling to identify low-level bottlenecks. The Codex-driven optimization loop—profile, suggest, deploy—offers a template for continuous improvement in production AI pipelines.
OpenAI API usage and cost
Running 10,000 agents for 36-48 hours consumes a substantial token budget. Third Moment reports using one trillion OpenAI tokens in a single day. While the article does not disclose pricing, the scale implies significant API costs, suggesting that enterprises must balance compute needs against budget constraints.
What’s next?
Third Moment plans to expand the agent ecosystem, integrating more specialized models and refining the synthetic-scenario generation pipeline. The next milestone will likely involve demonstrating hundreds of thousands of agents collaborating on a shared research task, a scenario that could expose new coordination challenges.
More from our coverage
For a deeper look at how enterprise deployments can transform internal workflows, see the Oracle ChatGPT Codex enterprise deployment Transforms Internal Workflows.
Related coverage
- TechCrunch Disrupt 2026 Offers $200 Ticket Discount as AI Leaders Converge
- Extend Builds Interactive Theranos Desk Simulator from Court Exhibits
- Samsung Tizen Update Policy Extends Smart TV Support to Seven Years