A hybrid state model for long-running agent work