sam-bench

Overview ¶

Command sam-bench measures what an agent experiences when it reaches the mesh through a sandbox boundary.

It issues a fixed workload rather than an interesting one. The chaos agent in this repo is driven by a real model and is the right tool for asking whether the mesh survives contact with an autonomous caller; it is the wrong tool for asking how long something takes, because it never asks twice for the same thing. This asks for exactly the same thing every time, so two runs differ only where the mesh differs.

One invocation is one observation: it records the workload, the latencies it saw, and what the processes involved reported about themselves before and after, into a single JSON document meant to be kept.