Methodology
A feat is a short card about one thing a real person got done with an AI agent: what the agent did, the concrete result, which agent, how strong the evidence is, and a link to the original post, repo, or article. This page says how a card gets made and what its labels mean.
Evidence grades
Every card carries one grade. It says how much proof the source gives, not whether the result is good. A card with grade C is not a bad card; it has less proof.
- Shown · grade A
An artifact is linked: a repo, a live site, a screenshot, or a video that shows the output.
Example: A GitHub repo with a README and a demo GIF.
- Numbers · grade B
A specific number is stated: time, money, a count, or a percentage, but nothing shows the output itself.
Example: “Twelve months of statements reconciled in one afternoon.”
- Claim · grade C
A claim only: the person says it worked, with no artifact and no number.
Example: “It set up my whole invoicing flow.”
A later reply with a repo or a screenshot upgrades a card to grade A after a person checks it.
Outcomes
Most cards are Completed, and then no outcome is shown. Partial means the agent did some of the job and a person did the rest. Failed, with a lesson means it did not work, and the source says what was learned.
How cards are chosen
A pipeline reads public sources every day: Hacker News, GitHub, YouTube, blogs and feeds, web search, posts on X, and submissions from readers. It keeps a post only when someone actually did something with an agent and says what happened. Opinions, announcements, ads, tutorials with no one doing the work, and vendor marketing with no named user are dropped.
A language model then writes the card from the source. Duplicates of the same story are merged, and the other links are kept as additional sources. Cards are ranked by evidence, how specific the result is, the kind of source, and how recent it is. Every card is then checked automatically: grade A or B, a documented result, no more than 25 words in a row taken from the source, a named agent, enough detail for its page, a high enough score, and no likely duplicate. A card that passes every check is published. A card a check flags waits for a person, who approves or skips it.
Workflow builders such as n8n, Make, and Zapier are out of scope. Other directories are never a source: every card links to the original.
Our own words
Every field on a card is written in our own words. No field repeats more than 25 consecutive words from its source, and an automated test checks every published field against the source text. The prompt on a card is written by Agentic Feats from the source; it is not the author's exact prompt. People are named by the handle they use in the source, never by a looked-up real name.
What search engines see
A card page is open to search engines when it has a result plus at least one more detail: the setup, a caveat, a prompt, or a note on the evidence. Thinner cards stay on the site but ask not to be indexed. Industry pages open up once they have enough feats from enough different people.
Corrections and removal
If a card is wrong, or you wrote the source and want it removed, use the takedown page. Removal happens within 7 days, with no questions. Agentic Feats has no relationship with any agent vendor; see about.
Open data
The open data page has the JSON and CSV files, the field list, and totals by agent, industry, and grade.