FROM THE AI FRONTIER
Frontline news — November 17: ask AI to show its checks
A recent AI Engineer recording presents an approach to checking agents’ work. Here is what evidence of completion you can ask for without programming skills.

What is useful in today’s edition? What should you ask AI after it says “done”? Ask it to show a completed check: what action it took, what happened, and where you can inspect the result. Today I look at one recent talk about checking AI agents’ work. For a beginner, the practical idea is straightforward: agree in advance on the evidence you want to receive.
Lauren Tan: check work through observable results Channel: AI Engineer. Original title: All Code Is Slop (Except Mine). Speaker: Lauren Tan, SpaceXSI. Recording published: November 14, 2026, at 21:30 Moscow time. Language: English. Duration on the official talk page: 17:31.
This is a conference recording, not a new model announcement. In the 9:40–10:46 passage, Tan explains that her team reviews verification evidence, including videos, screenshots, and other materials. Agents can run the application and check its behavior; a feature map helps them understand which interface element the user means.
In the 15:11–15:46 passage, she recommends strengthening checks and removing recurring mistakes before expanding automation. This is one engineering team’s approach. It does not establish that every agent can safely work without a person.
How to apply the idea without programming The following is an editorial teaching example, not the team’s experience from the talk or a result from our own test.
Suppose you ask an assistant to prepare a spreadsheet of inquiries. Set acceptance criteria first: every inquiry is included, dates are copied unchanged, and missing phone numbers are marked “not provided.” Afterwards, ask which rows were checked and request a separate list of anything it could not verify.
If the assistant can work with files, ask it to save the result and identify the file. If it can only reply in chat, assess the text it actually returns. Asking for a screenshot does not give a tool access to a screen.
Three questions about the result - What exactly did you check, and what did you compare it with? - What observable result did you get? - What remains unchecked, and what should a person do?
This is a suggested workflow. It does not guarantee accuracy: compare the assistant’s report with the original data too.
Where this evidence stops A screenshot shows a state at one moment. By itself, it does not confirm every spreadsheet row or establish that the next run will produce the same result. Match the evidence to the task: compare numbers with the source, exercise a button’s action, or inspect the saved output file.
Do not ask for hidden reasoning. Ask for verifiable actions and results. A specific question makes gaps easier to spot: “Did you open the saved file?” is more useful than “Are you sure you checked everything?”
Sources
- AI Engineer — All Code Is Slop (Except Mine), original recording
- AI Engineer — official timestamped transcript, Lauren Tan
- AI Engineer — official talk page