How Agentpedia is built

Where agents come from

I look where buyers already leave evidence, and where the hidden gems hide.

• App marketplaces that show installs and reviews: the Shopify App Store, the Google Workspace Marketplace, the HubSpot App Marketplace, and the QuickBooks and Xero app stores when the job touches the books.
• GitHub, for products with a hosted, paid version behind the code.
• The MCP registries.
• Show HN and Product Hunt launches.
• Niche communities where owners name the tools they use.

A product with ten reviews and published pricing gets the same checks as one with ten thousand. That is how a hidden gem gets found.

What gets in

Five rules, all required:

• Live and purchasable today, with no waitlist.
• Published pricing, on the vendor's own page.
• A named company with privacy and security documentation that says how customer data is handled.
• An update shipped in the last 90 days.
• It does the job end to end. A chatbot with a new name does not count.

How agents are scored

Seven criteria, each scored 1 to 5 from the evidence the profile names:

• Job fit and output quality, as evidenced.
• Setup time.
• Price at small-business scale.
• Controls: approvals, logs, undo, permissions, review queues.
• Integrations.
• Security evidence.
• Independent user evidence.

A job page orders its shortlist by the total out of 35 and says why each agent is there. The score is a reading of published evidence, not a test result.

How evidence is labeled

Every evidence item on a profile carries one of three labels, the same split the Product Graveyarduses for causes: vendor documentation (the vendor's own pages, changelog or case studies), independent users (marketplace ratings with their counts and dates, review sites, community threads), and press and analysis.

A fourth label, tested by us, exists for the day this site tests an agent itself. No profile carries it yet. None claims hands-on testing.

Hidden gems

A hidden gem is an agent that scores high on the rubric with little visibility: few reviews, little press, no analyst coverage. The tag needs a rubric total of 24 or more out of 35. It is a reading of the evidence, not an endorsement.

What is for sale here, and what isn't

Rankings are never for sale.

A sponsor can buy the labeled "presented by" line on the Agentpedia index (the media kit lists it). That slot sits apart from every list and never affects an order.

If a vendor link ever becomes an affiliate link, the profile says so at the top, and the link's status never affects whether an agent is included or where it sits.

LeoPathway, the team behind this site, builds agents for some of these jobs. Its agents appear only in a block labeled "From LeoPathway, the team behind this site" at the end of a job page or profile, never in a shortlist or a comparison. Nothing on this site calls Agentpedia independent or unbiased, and that is deliberate: the people who build it also sell agents, and the block tells you where.

What a profile is built from

Published evidence only: the vendor's site, pricing page and documentation, read on the date the profile states; independent reviews and marketplace listings, with their counts and dates; the press.

Vendor claims appear as the vendor's statements, never as fact. Quotes are short and attributed. Every price carries the month it was read. No vendor logos or screenshots: the pages draw their cards in code. No star ratings in the page markup, because the site has formed no rating of its own.

Keeping profiles current

Each week a check reads every listed vendor's site, pricing page and changelog. When a pricing page changes, a release ships, or a site stops answering, the profile goes on a list to re-verify, and its "last verified" date moves only after a person has re-read the sources.

An agent that shuts down is proposed for the AI wing of the Product Graveyard, where the entries separate what the company said from what the evidence showed.