Third Moment scales autonomous research with 10,000 AI agents
Third Moment Research says it is running 10,000 AI agents concurrently on autonomous investigations lasting 36 to 48 hours. The North American technology startup uses OpenAI’s Codex, GPT-6 Astra, GPT-6 Sol and GPT-6.1 Sol for research and computational work, and uses Daybreak to test the security of the environments in which its agents operate.
The company is building a self-improving system designed to examine agent behaviour, identify recurring weaknesses and create thousands of synthetic scenarios around those failures. Those scenarios are then used to improve agents, with the process repeated as new problems emerge.
Cutting backtest time and token waste
One early application is quantitative finance, where Third Moment can measure whether an agent’s work improves a trading system or produces noise. The team used Codex with GPT-5.6 to analyse thousands of recorded simulation trajectories, including agent actions, environment responses, timestamps and the research codebase.
That analysis found a bottleneck in backtesting. Running a trading strategy across a year of order-level US equity data took several hours. While the job was running, agents repeatedly polled for completion, consuming tokens and filling their context with status checks.
After profiling the backtest and examining inputs and usage, Codex identified a caching optimisation and a memory-packing technique. It tested solutions and deployed a fix to production. Third Moment says the backtest time fell from about three hours to roughly three to five minutes, a 97% reduction that left more token budget available for experiments.
Longer investigations inside controlled environments
For quantitative research, Third Moment runs the default Codex harness in a Dockerized environment. Each agent is provided with a terabyte of data, documentation and a research repository containing modules for feature engineering, GPU-supported model training, trading strategies and simulation.
GPT-6 Astra, GPT-6 Sol and GPT-6.1 Sol provide the reasoning layer, while the Codex harness lets agents use tools, create plans, set objectives, spawn subagents and compact context as work continues. When an agent stops before completing a task, Third Moment treats Codex as a subagent inside an outer harness powered by GPT-6 Sol or GPT-6.1 Sol. The outer system reviews the trajectory and can retry the work, end the job or send it for human review.
Third Moment says newer models can use experimental results to formulate follow-up work and investigate root causes through hundreds of parallel experiments. In one session, GPT-6.1 Sol identified the effect of non-displayed trades on an exchange where those trades represented an unexpectedly large share of volume.
Security testing before simulations launch
Scale also changes the security requirement. Before simulations begin, Third Moment uses Daybreak to automate tests of whether an agent could escape its research environment and reach the wider network. Adam Khakhar, co-founder of Third Moment Research, said autonomous penetration testing is necessary to test this infrastructure at scale.
The company plans to move from individual investigations towards hundreds of thousands of collaborating subagents, supported by chains of dependent experiments. For businesses evaluating autonomous-agent systems, Third Moment’s approach underlines the need to track execution time and token use, define escalation paths for incomplete work, and test isolation controls before expanding agent workloads.