So how exactly are people setting up these agents? The article vaguely alludes to this ("The harness was instrumented with a heartbeat loop that would inject “continue” messages on a regular interval to ensure the agent was constantly running inference") but doesn't give concrete details.

Is this literally just an infinite loop in a bash shell injecting the initial prompt into the OpenAI CLI, and each run of the CLI picks up where it left off using some kind of persistent memory? Or is it a single context window? It sounds like the latter but it's not clear to me how this "continue" message is "injected", and surely one context window would be inneffective after just an hour or two.

Sorry if this is a basic question but somehow I have missed the details of these kinds of agents.