An announcement doesn’t prove it works in real use.

Roundup articles will sell you eight company names and one big number. Where agents actually make sense already gave the test: a goal you can check, limited tools, a person who approves, and a way to undo. This series is how to read the stories that claim those jobs already work in real use.

Ask five questions, and write down the date with the number.

1. What could it touch?
2. What had to be approved?
3. Which number, on which date, from which URL?
4. Was a human still in the loop — to start, to send, or to merge?
5. What happened next?

If you cannot date the number, you do not have a result. You have marketing, not evidence. Why AI Agents Fail covers what happens when a claim was believed too early. What’s coming next names products as examples of patterns. This series is not that list.

Good prompts and review steps still matter. If a story says you can skip clear instructions or human approval, it is promising independence you cannot verify.

The series

  1. Support that takes action — Klarna’s two acts; vendor stories with their own numbers
  2. Operations with a copilot — a copilot that writes after a human starts
  3. Coding with human review — assign work, draft a PR, a person reviews