Let Agents In

Describe the problem in your own words, for example "let users upload images" or "send transactional email". Returns the vendors we have measured in that category, split by whether an unattended run clears every barrier we test, stops at one, or was never measurable, each with the date and a link to the evidence. This is not a recommendation: it does not know whether a vendor suits your job, only where an agent stops. Routing a sentence to a category is by far the weakest thing here, and the numbers below are the ones to plan around. Measured on 40 questions written by an agent with no access to this repository and no sight of the category list, labelled before the first run, and built to be hard: fourteen of them ask about the caller own code in commercial words (a billing module, a payments table, a notifications worker), and six ask for something real that a catalogue this size does not hold. It got 29 of the 40 right, said nothing on 8 it should have answered, sent 0 to the wrong category and answered 3 that it should have refused. Put another way: it gave an answer to 14 of the 40, and 3 of those answers were wrong, while refusing 18 of the 21 it should have refused. It is deliberately quiet. One vocabulary word inside a long question decides nothing, and a question shaped like a request for code rather than for a vendor is refused outright. Silence means we could not read the question, not that the category is empty. The tool reads English. Name the category yourself when you know it.

Scores a domain across 5 funnel stages using 16 deterministic HTTP checks. 18 points exist on paper; the score is out of the points that both applied to the domain and could be evaluated. Returns a per-check breakdown with the reason for each result and a permanent link.