Check out why Togerther.AI acquired CodeSandbox.

The obvious use-case (to me) is to create an agent that relies on an interpreter with a bunch of pre-loaded state that's already been set up exactly a certain way — where that state would require a lot of initial CPU time (resulting in seconds/minutes of additional time-to-first-response latency), if it was something that had to run as an "on boot" step on each agent invocation.

Compare/contrast: the Smalltalk software distribution model, where rather than shipping a VM + a bunch of code that gets bootstrapped into that VM every time you run it, you ship an application (or more like, a virtual appliance) as a VM with a snapshot process-memory image wherein the VM has already preloaded that code [and its runtime!] and is "fully ready" to execute that code with no further work. (Or maybe — in the case of server software — it's already executing that code!)

Main use case for me would be RLAIF. Given a prompt, generation, and a code execution result - rank N alternative executions and execution results for DPO/other training patterns.

In complex use cases like building a bi engineer, it’s useful to persist state across multiple function calls within the same interpreter.