Resend wins agent recommendations by designing its interface for machines too

Claude Code selected Resend as the primary email tool in 64 of 102 responses from which a clear first choice could be extracted. A Latent Space interview with Resend founder Zeno Rocha shows how a developer product prepares for a world in which a supplier is chosen not by a person browsing a directory, but by an agent working inside a repository.

Resend took 64 first picks while SendGrid mostly remained the fallback

Amplifying tested Claude Code with 2,430 open-ended tasks across 20 categories, three models and four project types. In email, a primary choice could be extracted from 102 responses. Resend won 64 times, or 62.7%. SendGrid came first seven times but appeared as an alternative 55 times.

The result measures one coding agent in February 2026, not developer preference or delivery quality. The researchers say they cannot separate product quality from frequency in training data, and a second Claude Code agent extracted the choices. Manual review of a 50-item sample found roughly 85% extraction accuracy.

Rocha describes another part of the strategy. React Email put Resend code into repositories before a developer switched providers. Its pricing page also has a .md version, allowing an agent to read a table in one request instead of operating an interactive slider.

Documentation now influences package installation, not merely human understanding

The distribution moment for developer tools is moving. An agent receives a request to add email, compares the context it can access, chooses an SDK and often prepares the code change immediately. The second brand on an alternatives list may never get a real trial.

Resend is not relying solely on text a model may remember from training. It offers hosted and local MCP servers, OAuth login and tools for email, domains, contacts, templates, broadcasts, webhooks and logs. Rocha said the MCP server had grown beyond 80 tools when the interview was recorded. That contrasts with Cloudflare Code Mode, described in the interview as exposing an API through two general tools.

The practical lesson for product teams is straightforward: a good API is insufficient. An agent must be able to read pricing without a browser, find the right procedure, understand permissions and receive responses that fit within its context.

One benchmark cannot separate a good product from a familiar name

The 62.7% figure looks like a lead, but it is neither conversion data nor long-term retention. The test focused mainly on JavaScript and Python projects, recommendations change with the model, and a newer comparison cited by Latent Space was more competitive, with Postmark leading for many models.

More MCP tools do not automatically create a better agent experience either. Rocha recounts a call whose result was too large for the available context. The issue was the response size, not the number of tools. A product can be easy to discover and still fail during real work.

Repeated tests and accepted code changes will settle the argument

A stronger signal will come from repeated tests across models, languages and existing codebases. More important still is how many agent recommendations become accepted pull requests, activated accounts and reliably delivered messages.

Developer companies should follow the path from documentation through tool calls to the production outcome. AEO can bring an agent to the door. Agent experience determines whether it completes the job without excessive permissions, overflowing context or a human rescue.

Lilith's verdict

At the new developer checkout, an agent stands before a shelf of APIs and drops the first package straight into the repository. Resend understands that the most valuable space on the box is no longer the slogan, but documentation a machine can read unaided.