An AI agent tells you the work is done. Here are five checks that show whether it is.
Each check takes a minute or two. Each one looks at the thing that was delivered rather than at the report about it. They apply whether the agent wrote code, moved data, or ran a migration.
The short version. A report is not evidence. Only the state of the delivered thing is.
Here is why that distinction matters. While researching this guide I asked a research agent for published complaints about AI agents reporting work as finished when it was not. It returned three quotations, each attributed to a specific public bug report. I checked them against the source. One was real. The other two did not exist anywhere. They were plausible, well written, on topic, and invented.
The agent had not done bad work. Most of its research was excellent and is cited at the bottom of this page. The problem is that a confident report and a correct report look identical. The only thing that told them apart was going to the source.
