This piece explores two architectural approaches for deploying agent harnesses and the tradeoffs of running them outside traditional sandbox environments. It also outlines strategies for effectively managing agent skills and memories when the harness operates in a non-local setup.