Definition
Plain language
Running a process thousands of times with random inputs to learn the shape of its outcomes.
As stated in the literature
A computational technique that samples random inputs from learned or assumed distributions to estimate properties of a system; in Agent JIT, used to choose execution strategies by simulating per-element latency draws.
Also called: Monte Carlo
Why it matters: It's the go-to way to reason about systems where the math is intractable but you can simulate the underlying randomness cheaply.
For example, to estimate how long a batch job will take, you draw thousands of random per-item latencies and add them up.
Heard on the show
“So they blend its estimates with full Monte Carlo returns from the actual trajectories.”Episode 008 — Why Long-Horizon AI Agents Get Stuck, and a Milestone-Based Fix That Helps