AgentsHow we source and rate
We cite only three kinds of sources: official vendor docs and pricing, independent benchmarks, and reputable press.
Rules
- Every price and feature links to an official page, with the date we checked it.
- Vendors can't pay for placement or rankings.
- We don't rely on social posts, anonymous reviews or affiliate listicles.
- Benchmark numbers belong to their publishers. We link to them rather than republishing them.
- Spotted a broken link or an outdated price? Email levelup@agentsthatwork.ai and we will fix it.
Benchmarks we trust
- SWE-bench Verified: Can a model fix real GitHub issues? The standard for coding agents.
- OSWorld: Can an agent complete open-ended tasks on a real computer?
- GAIA: General AI-assistant tasks needing reasoning, browsing and tools.
- HAL (Princeton): Reproduced agent evaluations, including GAIA.
- Artificial Analysis: Independent model intelligence, speed and price leaderboards.
Directory recipes
We gather agent recipes from many public sources: vendor template galleries, official examples, open-source projects, and builder communities on X, Reddit, YouTube and LinkedIn. A recipe is listed only if it is specific enough to follow and every app it connects to exists. We drop duplicates, spam, promotional listings and anything that relies on dead apps. The Verified mark means we found proof it works: a published template, an official example, or a public working demo. Community reports alone are listed without the mark. Licenses for third-party material are on our credits page.