What is actually known about the agents you can text

All 84 listings here state what they can do. 7 of 84 have ever been seen doing it in public, and 92 of the 98 posts on file belong to two products.

That gap is the most useful thing to understand about this market. It is not a criticism of any product in it. It is a fact about how new the category is, and it changes what a claim is worth.

Two records, and only one of them is big

This site keeps two kinds of record about every agent.

The first is what a company publishes. The capability rows, the price, the privacy policy, who carries the messages. All 84 listings state what they can do, 74 publish a privacy policy, 24 name the company that actually carries your texts. These are checkable and dated, and they are most of what this directory is.

The second is a sighting: a public post in which a named person says an agent did a specific job for them, filed with the post linked. There are 98 of those, and this is how they are spread.

The record is two products wide

7 of 84 listings have a single sighting. 77 have none.

Nothing about that is suspicious. Muse and Instinct are the products people were given first, and both have users who post. But it means almost everything written about what AI assistants can now do rests on two companies' customers choosing to tweet.

It also means the honest answer to "which of these actually works" is that for 77 of them, nobody outside the company has said in public. That is not the same as saying they do not work.

So the claims are the evidence

For most of this directory, what a product publishes is the only evidence there is. That is why the fields are recorded this hard, and why every one carries the date it was read.

A capability row is a claim. A price is a claim you can hold a company to. A privacy policy is a document with legal weight. None of them is a demonstration.

The 84 listings were read for nine things, each one a sentence a company could publish today at no cost. This is how many answer each.

Almost everyone says how you start, and almost everyone has a policy. The questions that get dodged are the ones with a named third party on the other end: who carries your messages, whose models read them, where to send a security report.

The market is not split into the open and the silent

A reasonable guess is that this divides into companies that publish and companies that do not. It does not.

There is one hump, not two. Most listings sit in the middle, answering five, six or seven of the nine, which means disclosure here is not a stance a company takes. It is a set of small decisions, each one made or skipped, and almost every product has made some and skipped others.

What this page will not do

It will not rank them. The market map places all 84 on two recorded fields, and says in its own copy that the rank correlation between going further and publishing more is 0.12. That is nothing. A product that does a great deal and says little is not thereby worse than one that does little and says a lot.

It will not treat a company's account of itself as evidence of what it can do. When a founder describes their own product, that is recorded as the company's words and moves no field here.

And it will not fill a blank with an estimate. A third of this market publishes no price, most of it has never been seen working by a stranger, and both are findings rather than gaps in our research.

What to watch

Whether the products anybody has posted about passes ten, which would be the first sign the public record is broadening rather than deepening around two names.

And whether the share publishing no price moves at all, now that one funded company has published one and left that set.