Building what's next (Interview)

Zeno Rocha, founder and CEO of Resend, explains how AI agents misread his company's pricing page because they "cannot drag a pricing slider," leading him to serve a markdown version through content negotiation. He argues that building for agents means following established standards - RFC rate-limit headers, publishing the Open API spec as JSON in the.well-known folder - rather than inventing proprietary formats. Agents, he notes, "couldn't care less" about animated icons or keyboard shortcuts that matter for human developer experience.

Rocha walks through Resend's architecture: starting with a TypeScript/Node stack on AWS, then rewriting the email-sending pipe in Go to handle concurrency when customers send multi-million-email broadcasts simultaneously. He describes the company's growth to "three million users total 100 000 paying customers," the ongoing battle against spam gangs exploiting the free tier, and a shift in the user base from traditional developers to "builders" using tools like Lovable and Bolt.

The conversation frames agent experience as a distinct discipline from developer experience. Rocha describes a closed-loop system where down-votes on Resend's in-app agent trigger draft PRs that update tool calls and evals, and he emphasizes that error messages, CLI output shapes, and granular API permissions are the primitives that determine whether an agent can self-correct or gets stuck.

• What if you served a machine-readable version of your pricing page to agents?

  • Move: Add a markdown or JSON endpoint for your pricing page, served via content negotiation (e.g., Accept: text/markdown), so agents can read exact tiers and costs without interacting with UI sliders.
  • Why Now? Rocha discovered that both Claude and ChatGPT misidentified Resend's pricing because agents "cannot drag a pricing slider." As agent-driven comparison and purchasing grow, inaccurate pricing signals will misroute potential customers away from your product.
  • Expected Upside: Agents can correctly recommend your product in comparisons, reducing lost signups from false "most expensive" labels. This is a low-effort, high-leverage fix that Rocha implemented in a single Saturday afternoon.

• What if you followed the RFC for rate-limit headers instead of inventing your own?

  • Move: Implement the three standard rate-limit headers (remaining count, time unit, retry-after) on every API response, then test that an agent can auto-retry without human intervention.
  • Why Now? Rocha notes "if you follow the rfc agents will know they will know how to auto retry." Proprietary headers leave agents stuck, forcing them to escalate to a human operator and breaking unattended automation loops.
  • Expected Upside: Agents handle throttling autonomously, reducing support tickets and enabling longer unattended workflows. This is a proposal grounded in the source's observation that standardized primitives are "the most important thing that matters right now."

• What if you published your Open API spec as JSON in the.well-known directory?

  • Move: Add a CI step that converts your YAML Open API spec to JSON and deploys it to /.well-known/openapi.json so agent tools can discover your API without manual configuration.
  • Why Now? Rocha found that agent-experience scoring tools were "searching for the dot json version of the open api spec Not the dot yaml," and his own agent couldn't locate the spec until he added the JSON endpoint.
  • Expected Upside: Agents can self-discover your API surface, reducing onboarding friction for agent-driven integrations. This is a practical, one-time infrastructure change with compounding discoverability benefits for every agent that encounters your service.

• Serve a markdown or JSON version of your pricing page via content negotiation so agents can read exact costs without UI interaction; Rocha fixed a mispricing issue this way in one afternoon.
• Implement RFC-standard rate-limit headers (count, unit, retry-after) on all API responses so agents can auto-retry without escalating to a human operator.
• Publish your Open API spec as JSON in the.well-known folder and add a CI job to keep it in sync with your YAML source, because agent tools search for the JSON variant specifically.
• Design your CLI with a non-interactive mode for every command and flat JSON output (not nested arrays) so agents can pipe results through jq without parsing confusion.
• Write specific, actionable error messages that tell an agent what went wrong and what the next step is, rather than generic codes that leave it stuck; Rocha is "revisiting the error messages now because the error messages are crucial."