Designing for Agents — Zeno Rocha, Resend
- Source
- latent.space
- Author
- swyx .ai
- Date

Context
Latent Space's episode with Zeno Rocha, co-founder and CEO of the email-sending company Resend, which he says had 4.3 million users and 120,000 paying customers when the episode was recorded, starts from a puzzle: why do coding agents keep recommending Resend even to a team already on Cloudflare, which has its own email service, and over much larger incumbents? The episode's written introduction cites a study in which Claude Code from Anthropic named Resend first in 64 of the 102 email recommendations the researchers could extract, while SendGrid came first seven times and appeared 55 times as an alternative. The introduction's point is that the first pick matters: a developer who asks an to add email and accepts the patch may never get around to trying the second choice. Latent Space adds that its own, more comprehensive tracker of what AI models recommend was more competitive, with Postmark on top for most models.
The introduction points to choices that put Resend in front of both people and agents. Its first useful distribution, the introduction says, came from React Email, a library for building email templates. Resend documented how to send those templates through SendGrid, so some of Resend's code was already in a project by the time its developers wanted a different provider. Resend's pricing page has a slider for people, and adding .md to the page's URL gives an agent a plain table it can read in one request. Resend's server, which lets an agent operate the service from a chat client, grew from about 20 tools at launch to more than 80 so that users can do there what they can do in the dashboard. The write-up contrasts that with Cloudflare's server, which exposes its API through two tools, and notes that a smaller tool count would not have shrunk the oversized tool result Rocha found when he opened his own sessions. Its suggestion for MCP designers is to look at the full result of a call they make every day and see how much of it the model needs.
Everything above comes from the hosts' written introduction to the episode, not the recorded conversation. Deliverability questions the page only previews are argued in the audio: for example, a provider that rewrites links through a shared click-tracking domain can see that domain blocklisted when one customer uses it for phishing, which hurts other customers' legitimate mail, and Rocha explains why Resend introduced custom tracking domains.
- In an study cited by Latent Space, Claude Code picked Resend first in 64 of 102 extracted email recommendations. SendGrid came first only seven times but appeared 55 times as an alternative, a gap that matters when developers accept the agent's first patch.
- Resend's first useful distribution was React Email, a template library that also worked with other senders such as SendGrid, so Resend code was often already in a repository by the time a developer went looking for a new email provider.
- Resend's pricing page uses a slider for people, but adding .md to the URL returns a table an agent can read in one request. The post suggests checking whether a model can answer a customer's pricing question from your own page.
- Resend's MCP server grew from about 20 tools to more than 80, the opposite of Cloudflare's two-tool search()/execute() Code Mode design. In one of Zeno Rocha's sessions a single result was too large for , which the post notes fewer tools would not fix.
- Shared click-tracking domains let one customer's phishing get the domain blocklisted and damage links in other customers' legitimate mail, the problem behind Resend's move to custom tracking domains, according to the episode notes.
- AI agent — An AI system that doesn't just answer once but works toward a goal in a loop — taking actions, reading the results, and deciding what to do next.
- MCP — The Model Context Protocol — an open standard that lets any AI assistant plug into any tool or data source without custom integration code.
- generative engine optimization — Making a site the source an AI assistant cites when answering, rather than a link in a list of ten blue results.
- context window — The maximum amount of text a model can consider at once — its working memory for the current conversation or task.
article[AINews] Fable 5.2 retakes World #1 LLM spot as $ANT IPO roadshow begins
articleWhat Frontier Labs Buy: A Data Deep Dive with Handshake AI — Garrett Lord & Team
articleHow Stanford Teaches AI-Native Engineering — Mihail Eric, Monaco
article[AINews] Inkling 2 Small: 17A-594B trained in 68 hours on 12T tokens, beating Muse Spark
Checking sign-in…
Loading comments…


