Oh I did read the README, but still have the question: while it does save on cost, latency and complexity, the tradeoff is that the agents can't run whatever they want in a sandbox, which would make them less capable too.