The former. I would not expect any candidate - especially a new grad - to have actually used GasTown style setups unless they were coming from another company with money to burn on those sorts of multiple-agent scenarios. I barely run more than one or two Claude sessions at once because it slams into the usage limit so fast.
I could steer the conversation to explore if the candidate had insight into why GasTown or other "swarm" models exist though. Do they know what a system prompt is? How context is managed/affects output? How harnesses can delegate to simpler models, and when to use different classes of model? I wouldn't expect anything particularly deep here, but bonus points for a solid understanding/experience of the tooling and how to get the most out of token budgets is valuable.
To the parent point: even if you refuse to use LLMs for whatever reason, I would expect any coding applicant to have at least tried them. I'd assume the person is living under a rock if they said no.
I could steer the conversation to explore if the candidate had insight into why GasTown or other "swarm" models exist though. Do they know what a system prompt is? How context is managed/affects output? How harnesses can delegate to simpler models, and when to use different classes of model? I wouldn't expect anything particularly deep here, but bonus points for a solid understanding/experience of the tooling and how to get the most out of token budgets is valuable.
To the parent point: even if you refuse to use LLMs for whatever reason, I would expect any coding applicant to have at least tried them. I'd assume the person is living under a rock if they said no.