Two architectures for running agent harnesses, the tradeoffs between them, and how we make skills and memories work when the harness isn't local.