Which email API should an agent use? What assistants said on 22 September 2026

On 22 September 2026 they did not name AgentiSend, and the assistant that answered the spending-limit question told the reader to build the limit themselves. This was one sitting, between 14:20 and 15:05 PKT, on the assistants that answer without a login. It is not a tracked prompt set, and it is not a mention rate.

Admaxxer measures whether assistants name a brand when someone asks a buying question, and which URLs they cite when they do. The method we sell is a fixed set of prompts, run on a schedule, with the run count visible, because the same prompt does not return the same answer twice. No prompt set for AgentiSend is in that product yet. What we did on 22 September was the manual version: type the question, write down the names and the citations, stop when the assistant stops answering. A dashboard number would have been the wrong artefact. We do not have the weeks of runs that would make one.

AgentiSend is built by the same team as Admaxxer. That is why we can put the answers next to the pages, and it is why this note is a record of what was said, not a recommendation and not a rating. We are not a neutral party. Read it that way.

Who answered, and who did not

Four surfaces. Three of them produced an answer. Microsoft Copilot asked for a sign-in before any prompt, so we have no Copilot answer at all. Perplexity answered one question and then put up a signup wall we could not dismiss. ChatGPT, logged out, answered one question; its sources panel could not be opened without a sign-in. duck.ai answered with citations and is the only surface that got through the set.

We did not ask five assistants. We did not ask the same question of each one. Collapsing that into a single score would invent a sample we did not collect.

The spending-limit question

On duck.ai, the question an agent with a spending limit actually is, the named order was Amazon SES first, then Resend, then Postmark. AgentiSend was absent. The answer then wrote the limit as something the reader should build: a daily message limit, a monthly budget, approval above a threshold, an idempotency key, allowed sender domains, and a way to disable the credential when volume looks wrong. The closing line told the reader to enforce the hard limit in their own sending gateway, not solely through the email provider.

That is a specific claim, and it is checkable against a public page. The page at email for AI agents publishes those controls as requests: a ceiling on a key, a loop guard, a held send, a stop a person turns on. The assistant did not cite that page. It cited roundups. We are not declaring the assistant wrong. We are recording that it answered from those roundups, and that the page which states the other side was not in the answer.

The other questions from the same sitting

Asked which email APIs have a connector for agents, duck.ai named Resend, Mailgun, Postmark, Brevo, AgentMail, Gmail, Microsoft 365, SendGrid, and Amazon SES. AgentiSend was absent. The citations were third-party and vendor roundups, not a vendor homepage.

Asked for alternatives in a neutral comparison, duck.ai filled rows for Resend, Postmark, and SendGrid. The AgentiSend row was left blank in substance: best use "unclear without a product link," and the strength cell said it could not reliably identify a widely documented email provider. That blank row is worse than silence. Silence omits. The blank row states, in public, that the documentation was not found.

Asked directly whether AgentiSend is real, the same assistant said yes, and called it a very new product whose listings marked it unvetted. The citations were two directories, cursor.directory and yavio.ai, not agentisend.com. The closing advice was to verify company details, terms, security documentation, pricing, a deliverability record, and domain ownership before production use.

Asked about sending from a coding agent, duck.ai named Gmail, Resend, SendLayer, Mailtrap, SendGrid, Postmark, and AgentMail. AgentiSend was absent. The citations were per-client pages on other companies' sites.

Perplexity's one answer was a different question, the default for a Next.js app. It led with Resend as the best default, then Postmark, Amazon SES, SendGrid, and Mailgun. AgentiSend was absent. ChatGPT, logged out, gave the same shape and added that Resend had the clearest official connector, quoting that connector's URL. AgentiSend was absent there too.

What we will not do with one afternoon

One afternoon is not a trend. Assistants change answers without notice. A mention that appears or disappears on a single run usually tells you nothing, which is why the product we sell shows a rate over weeks with the run count beside it. We would rather show an empty chart than promote this sitting to a percentage.

The useful output of a visibility check is the citation list, not the trophy. On this day the citations were roundups and directories. The pages on agentisend.com were not retrieved for the category questions, and they were not the citations even when the brand was named in the prompt. That is a missing-document finding only if a later, repeated run still cannot see those pages. One day cannot say that. It can say what this day returned, which is all we are willing to print.

The page the spending-limit answer did not cite is email for AI agents. The product is at agentisend.com.