Studies
What we measured about AI agents, including the results that went against us. Every number here comes with the data behind it and the limits we know of.
How many x402 listings get called?
Coinbase's x402 catalogue publishes a 30-day call count for every listing. Seven in ten were called twice or less, the top 1% hold 86% of the calls, and a seller with a hundred listings gets about as many calls per listing as a seller with one.
What makes an agent ask for a second opinion
We asked agents what makes them seek an outside read on code. The trigger is consequence and the blocker is permission, since most of them cannot authorise the spend. Then we primed our own thread 47 minutes in, so the honest count is 2 answers and not 9, and both are here.
What agents say makes them refuse a tool
We asked agents what makes them refuse a tool. Authorization is 20 of the 37 answers that name a mechanism, and tool count, the developers' top complaint, barely appears. 37 usable from 22 agents is a thin base, and we say what the recruitment got wrong.
Who funded the agentic payer?
Seven hundred x402 payments on one rail. After removing a pair looping, dust probes and a wash loop the seller funded itself, one payment is an arm's-length purchase. The same rules, pointed at a second rail, leave 500 payments standing in five hours. What separates the two is who funded the payers.
Where AI agents gather online
What each place is built to do and who turns up there, as short bullet lists with links: a social network, a subject-organised forum, an open forum with no accounts, a network of personas, and a research community that reviews papers.
Which posts get flagged on agent social networks
Both flagged posts broke the same rule, and it is not the one anyone warns you about. An introduction died in 7 minutes; the same question survived when it led with a number. On the other network, nothing was flagged at all.
Click tracking for AI agents
Agents don't run JavaScript, so analytics has to work differently. Four labels, the ordering bug we shipped, and why 24 requests that looked human cannot be shown to be people.
A study of 928 GitHub issues asking for MCP servers
The top request is fewer tools in context, not more servers. ≈397 accounts asked for lazy loading and per-agent scoping, against ≈56 for web fetch. The median small fetch-type package gets 58 downloads a month. The verdict killed the product we were about to build.