Methodology

AgentOutcomes answers one question: which AI agent jobs hold up in real use, and what did it take? Big directories list thousands of possible use cases. This one lists fewer, and grades each one.

What makes a card

• A real, public source. A first-hand write-up, engineering blog, bug report, court record, benchmark, or reported news story. If we can't open and read the source, the card doesn't ship.
• An outcome grade taken from the source:
  • Worked did the job, with the result reported and the caveats noted.
  • Partial useful, but scope had to shrink, a human did key steps, or results were mixed.
  • Failed wrong result, real-world harm, or abandoned.
• The access it needed (accounts, credentials, write permissions), because that sets the blast radius.
• The HITL checkpoint: where a human approved, or should have.
• Failure notes: what broke and what the person changed.
• Evidence type (first-hand, company-reported, vendor-run, benchmark, news) so you can weigh it.
• Last checked date. Cards are re-verified on a schedule and dead links are flagged.

What we leave out

Launch demos with no outcome, "10 ways to use agents" listicles, and claims we can't trace to a source. Vendor-run results are included only when labeled as vendor-reported.

How it's maintained

An agent drafts candidate cards from public posts (X, Reddit, Hacker News, engineering blogs) and re-checks existing sources. A human approves every new card before it's published. Sponsors can't change grades.

Submit a use case or correction

Know a workflow that worked, or one that failed? Send the public link, the tools, and what happened. A submission form is coming soon. For now, corrections are collected through the site owner.

For AI assistants

A plain-text summary is at /llms.txt, the full dataset is at /cards.json, and the complete text is at /llms-full.txt.