
cpu are fast, but every agent needs a little sandbox. agents doing api/browser requests is one thing. but agents tackling c++ projects need to build all that. agents working on your mobile/web application need to spin it off, click around in a simulator/browser.. with all the interleaved thinking and tool use those environments must persist during the whole agent workflow. sitting idle and waiting for model to decide what to do next. just think that you are adding way more developers and each of them needs a compute. all that compute demand is new.









