How we verify

Every example on Real Agent Cases links to where it came from, so you can check it yourself. Here's what we do before listing one.

• We read the original source.Before we list an example, we open the original post, video or article and read it.
• We only list what the source shows.If the source doesn't back up what happened, we don't list it.
• We label every prompt."Original prompt" means the exact words the person used. "Recreated prompt" means they didn't share their exact words, so we wrote one based on what they described. On failure and partial cases, an "Improved prompt" is our stronger version, written to avoid what went wrong. It hasn't been tested by the poster, and the prompt they tried is shown under "What the poster tried".
• We include failures.If the agent only got part of the way, or didn't manage it at all, we say so.
• We show two dates.The source date is when the person posted it. The last-verified date is when we last checked it.
• Numbers are their claims.When someone says they saved 10 hours or $500, we report that as what they said, not as something we measured.

Evidence labels

Each example has a label that tells you what kind of proof the source gives. When we're not sure, we use the more modest label.

• Evidence: Shown on videoThe source includes a video or screen recording of the agent doing the task.
• Evidence: Screenshots or logsThe source shows screenshots, a transcript or the agent's output.
• Evidence: Described by posterThe person describes what happened in words, without screenshots or video.
• Evidence: Company case studyA write-up published by the company that makes the agent or uses it.
• Evidence: Article or interviewA newsletter, journalist or interview account of what happened.

Failures and partial results

We keep cases that went wrong, because they show where agents still struggle. We mark a case Failure when the source says the agent didn't get the job done, and Partial when it only partly worked or had a significant problem a person had to deal with, like wrong output, a step it couldn't finish, or work the person had to redo by hand. Minor caveats don't count. Each tagged case explains what went wrong, in our words, based on the source.

Beginner-friendly

We mark an example Beginner-friendly only when the task needs no logins or account access, no payments or purchases, and no coding or terminal, and it worked. You should be able to paste the prompt and get something useful, with little risk.

Spotted a mistake or a broken link? We'll take a look once submissions reopen.