How it works
Set up a Docker environment per agent with the data, documentation and a research repository with modules for features, model training, strategies and simulation.
Run the default Codex harness inside that environment so the agent can plan, set objectives, spawn subagents and compact context across 36- to 48-hour sessions.
Give each agent a baseline system to improve and have it test strategies on historical data and judge whether each gain is real or noise before choosing the next experiment.
Treat Codex as a subagent of an outer harness that, when a task is left unfinished, examines the trajectory and decides whether to retry, terminate or escalate to a human.
Use Codex to analyze recorded trajectories, timestamps and the codebase to find bottlenecks in the environment itself that waste tokens.
Before launching simulations, test whether an agent could break out of its environment and reach the network (Third Moment uses Daybreak for this).
Prompt
Replace anything in [brackets] with your own details.
Analyze the recorded trajectories of our research agents (actions, environment responses and timestamps) together with the repository code. Find where agents lose the most time or tokens, for example waiting on long-running processes or repeating status checks. Profile the slowest step, propose optimizations, test each one and measure the time saved without changing the results. Deliver the fix with tests and a before-and-after summary.Reported result
Codex found a caching and memory-packing optimization that cut the backtest from about 3 hours to 3 to 5 minutes (-97%). The company reached 10,000 concurrent agents and 1 trillion tokens in a single day (figures from OpenAI).
As reported by the source. Results depend on your setup and your data.