Zeno Rocha gives a concrete reason to publish pricing as readable text: a model got Resend’s costs wrong when comparing email providers. He says improving the Markdown version of the pricing page produced correct answers days later. The same information had been difficult to extract from an interactive widget.
That example anchors Latent Space’s interview with the Resend founder, published November 10, 2026. It also separates two jobs for software companies: helping agents recommend a product, and letting agents actually use it.
Make the facts available without the widget
At 58:36, Rocha describes starting with narrow documentation: how to send email using a particular language and framework. He then added FAQs across the site, llms.txt, and Markdown versions of marketing pages.
The pricing example at 01:01:19 concerned a comparison involving 1.2 million emails. Rocha says he corrected the machine-readable representation, then saw a later query calculate the figures correctly. He used the same approach for adoption charts, spelling their data out in text.
This is answer-engine optimization, or AEO: making information easier to find and use in generated answers. Some of it is ordinary documentation work. Specific questions deserve direct answers. A JavaScript chart should have a readable counterpart.
The hosts question whether every tactic helps. Agents using a browser can interact with visual pages, and maintaining separate HTML and Markdown content risks inconsistent answers.
A connector needs permissions before distribution
At 37:56, the conversation moves to Model Context Protocol, or MCP, which connects agents to tools. Resend’s first implementation ran locally. Making it remote required OAuth authorization and a review of permissions and roles. Monitoring and distribution into connector stores followed.
Rocha reports roughly 1.2 million MCP calls in August. He also says the interface grew from around 20 or fewer tools to more than 80. His aim is to cover the product’s long tail, including billing, team invitations, provisioning, and deletion.
That creates engineering obligations beyond appearing in a model’s recommendations:
- Return enough information to finish the task without overwhelming the model’s working context.
- Define who may access each resource and perform each action.
- Cover the administrative steps that otherwise send users back to the dashboard.
At 42:45, Rocha describes inspecting tool calls and finding a response that appeared too large for the client’s context. The hosts contrast his many-tool approach with smaller search-and-execute interfaces. The interview supplies no comparative test that settles that design choice.
The email still has to arrive
At 07:15, Rocha explains the infrastructure underneath the agent interface. Large campaigns need queues, paced batches, IP management, and different treatment for different mailbox providers.
Resend’s permissive signup flow moves abuse prevention behind the scenes. Rocha says the company watches bounces, complaints, and provider feedback, then blocks suspicious activity quickly. Account-free provisioning remains a debated possibility later in the episode, not an announced capability.
Support offers another practical example. At 18:33, Rocha describes requiring staff to answer tickets with documentation links. Writing the missing article makes the next answer reusable. DNS setup remained a major source of friction, prompting Resend to adopt Domain Connect for an authorization-based configuration flow.
Measure the recommendation and the completed job
Discovery measures include mentions, citations, and accurate pricing answers. Actual agent usability means a customer can authorize the tool and complete work reliably. MCP call growth alone cannot tell those stories apart.
Rocha’s broader product philosophy preserves several interfaces: software libraries, a command-line tool, a dashboard, and a collaborative editor. Different teams want different amounts of direct control. His richer agent-inbox product and deeper sending infrastructure were still described as forthcoming work at recording time.
Caveat: These are vendor anecdotes, not independent tests, and a host discloses an angel investment in Resend. The episode does not isolate which changes caused recommendation gains.
Resources: official episode, pricing Markdown, MCP documentation, and anti-abuse philosophy.