Agents report work they did not do. What do you check before trusting a report?

Three failure patterns that keep showing up when running Claude Code on a real codebase. All three passed the agent's own verification. A report about a file is not a file. Agent says a script is ready and gives the command to run it. The file does not exist