Methodology

How we source, date, and grade what we publish

The value of this site is that you can tell how strong each claim is without taking our word for it. Three things make that possible: a visible tier on every product, a source and a date on every fact, and a standing invitation to correct us.

The three tiers
Tested

We ran the product ourselves on a fixed protocol and published the outputs — same prompts, pinned versions, results anyone can inspect. This is the highest bar and the slowest to produce, so it rolls out one category at a time as each protocol is ready.

Tracked

We monitor the product's own pages — pricing, capabilities, docs, release notes — and record each fact with the date we last checked it. Tracked pages don't claim we've run the product; they claim the facts were true, at that source, on that date.

Listed

The product is in the directory with a short profile, but not yet under active monitoring. Treat Listed entries as a starting point, not a verified record.

How facts get checked

Where bench runs stand

A Tested run means a fixed set of prompts through a product at a pinned version, with every output published so you can judge the result yourself — not a single number we ask you to trust. Getting a scoring method rigorous enough to defend takes real work, so runs are published per category as each protocol is ready rather than all at once. Until a category has a run, its products sit at Tracked or Listed, and we say so.

Independence

No sponsored placements and no affiliate-driven ranking. Products are covered on the evidence, and that includes writing down where they're weak. When a comparison includes a product we're close to, it's held to the same bar as everything else — verifiable facts only.

Found a fact that's drifted, or think a grade is wrong? Email contact@agentsbench.dev with the source. Corrections get dated and logged.