After the Recommendation: What Resend’s Agent Strategy Requires Beneath the Interface

1. Guest Background

Zeno Rocha enters this conversation with experience on both sides of a developer tool: as an open-source creator concerned with the experience of using software, and as the founder responsible for the infrastructure underneath it. He founded Resend and created the Dracula theme. In this approximately 86-minute Latent Space interview, hosts Alessio Fanelli and Shawn swyx Wang explore how those interests now meet the demands of AI agents.

Rocha describes himself as a frontend developer who had never built an email pipeline before starting the company. His opening thesis concerned an underserved developer audience within a large, established market. Imagining a Stripe for email gave him a product direction before he understood the operational depth of delivery.

His account of access is equally personal. He connects Resend’s free tier and investment in multiple language SDKs to growing up in Brazil with a poor computer and internet connection. After finding generated SDKs unsatisfactory, the team chose to write multiple versions by hand. That explains his priorities; it does not establish the return on every investment.

The interview also puts ambition in context. Rocha recalls joining YC Winter 2023 with a roughly one-year-old daughter, initially wanting to lead his batch, then coming to see entrepreneurship as an ongoing process. His hypothetical comparison between a close family relationship and a larger company expresses a preference, while his admission of continuing conflict between family and executive roles keeps that preference from becoming a tidy success prescription.

2. What the Episode Covers

The episode examines what an email platform must become when customers increasingly begin their work inside an agent. Rocha says Resend was serving about 4.3 million users on the recording day, with customers including Warner Brothers, HBO and independent developers. These are his reported figures, providing context for a discussion that moves between individual onboarding, large sending operations and enterprise requirements.

React Email illustrates the company’s approach to finding an entry point. Rocha corrects the suggestion that the open-source project preceded the business: Resend came first, and user research subsequently exposed the difficulty of building email templates. He deliberately wrote integration documentation for competing sending platforms, including SendGrid, Mailchimp and Mailgun. The template tool could therefore solve a useful problem before asking someone to choose Resend for delivery.

Its technical scope matters. React Email uses JSX templating and compilation to handle email’s tables, inline CSS and Outlook-specific compatibility work, with TypeScript and Tailwind tokens available to authors. The interview explicitly distinguishes that workflow from shipping React state or JavaScript interactions inside an email. Familiar authoring tools simplify a constrained output format; they do not remove its constraints.

The infrastructure story develops in the opposite direction, from a higher-level experience toward more underlying control. Resend initially used AWS SES while concentrating on SDKs and developer experience. Rocha describes limitations around domain organization and configuration sets, alongside the surrounding work required for events, logs, queues and webhooks. He says the company is now building IP-pool management and provider-specific traffic control, maintaining mailbox-provider relationships, and optimizing p99 latency. The sending path has moved from JavaScript to Go. Better delivery and visibility are stated objectives, without quantified improvements in the interview.

Rocha calls the broader product ambition formless: its capabilities should be accessible wherever work starts. He reports that MCP usage later at least doubled monthly and reached roughly 1.2 million calls in the August discussed. Those numbers describe activity, rather than demonstrated task quality.

Multiple interfaces also serve different human workflows. Marketing teams can own visual templates while engineers supply a template ID and variables; other teams want templates under source control. The choice reflects who owns template changes within a particular team. Rocha describes about 20 colleagues jointly editing an all-hands email. Alongside this, he previews programmable inboxes with creation, reading, read-state, starring and labeling APIs. That product was still forthcoming. The hosts describe a roughly 50-person distributed company, while Rocha says enterprise selling is only beginning and emphasizes that enterprise customers require a different product, beyond adding identity and audit features.

3. Core Views: Reasoning, Examples, and Limits

A central argument emerges from the gap between accepting a request and achieving the customer’s purpose. In email, raw sending capacity leaves much of that gap unresolved. Rocha explains that Gmail, Outlook and Yahoo Mail behave differently, requiring queues, batches and provider-specific pacing. In his example of sending 100,000 messages, early recipient engagement can influence what happens to later messages, making audience ordering relevant. IP-history checks and pool management add another operational layer. His account explains why more compute alone cannot settle delivery, although the interview does not independently test the mailbox mechanisms he describes.

Removing onboarding friction transfers responsibility to the platform. Resend’s approach, Rocha says, is to trust new users initially, block quickly when bounce rates, complaints or provider feedback indicate a problem, then investigate through conversation. He describes a Japanese abuse group bringing roughly 20,000 domains and identifies signals such as very new domains, lookalike names, missing normal websites and abrupt sending bursts. These examples show the detection burden created by openness. They also reveal a boundary: a suspicious signal is not proof of intent. Rocha says most cases are misuse rather than abuse, so explanation and recovery belong in the operating model alongside enforcement.

Remote MCP exposes a comparable gap in product architecture. The first local implementation came from Yoko at a16z. Turning it into an official remote service required OAuth, a review of resource permissions, monitoring and efforts to reach official marketplaces. The work was therefore about identity and operational ownership as much as tool definitions. The host’s proposal for a headless backend serving multiple clients follows naturally, but remains an architectural position expressed in the discussion. Rocha supplies a concrete maintenance example: private dashboard APIs and public APIs created duplicated code; a monorepo improved reuse without eliminating the burden of maintaining both.

Feature counts provide an incomplete answer to whether that architecture works for agents. Rocha says Resend’s MCP grew from around 20 tools or fewer to more than 80. His attention has shifted toward oversized responses and difficult operations such as billing, team invitations and account creation or deletion. His investor update, assembled using MCP, supplies a real usage example. It does not supply a systematic completion rate. The same distinction applies to distribution statistics: he reports SDK npm downloads above twice SendGrid’s while acknowledging that SendGrid sends more email. The host’s claim of recommendations in more than 60 percent of Claude Code cases comes without a sample or method. Downloads, recommendations and calls measure different stages, none of which independently proves market share or reliable execution.

The AEO discussion makes the evidentiary boundary especially clear. Framework-specific documentation, FAQs, llms.txt, Markdown pages and textual chart data address whether a model can obtain useful information. Rocha’s pricing example addresses a narrower outcome: a model incorrectly calculated the cost of 1.2 million emails, then answered correctly several days after he supplemented the Markdown page. That is an informative anecdote about answer accuracy. It does not establish increased recommendations or isolate the content change as the sole cause. The host explicitly questions distribution effects, and Rocha acknowledges the cost of keeping web and Markdown versions aligned.

Two closing themes reinforce this resistance to easy proxies. Rocha treats design and product identity as differentiation, while admitting that polished fundraising slides could consume misplaced effort. No causal growth result is established. The host likewise argues that next-generation or AI-native SendGrid supplies a reference point without yet defining a new category. Operational comparisons need equal care: marketing delivery involves changing batches, intervals and domain-specific rates. Rocha’s account of taking more than a day to email the stated 4.3 million users cannot be combined with the host’s hypothetical 100,000 messages in ten minutes into a performance benchmark. His concern about contact-based pricing for rapidly growing customers further separates pricing units, audience size and sending behavior.

4. Learning and Application

For a developer-tool team, the practical starting point is a complete customer task. Use the gaps Rocha identifies to examine a workflow that includes reading usage, resolving billing, inviting a colleague or closing an account. Check resource authorization and response size at each step. If a user must return to the dashboard for a decisive action, describe that boundary explicitly. This proposed evaluation follows from the episode’s examples: tool count should help explain available capabilities, while a completed workflow establishes whether those capabilities fit together. Expanding execution requires permission design alongside endpoint coverage.

Support offers a second application: move from answering a recurring question to making its answer reusable, then consider removing the cause. Resend requires replies to link to documentation, with a new article written when necessary. Rocha’s example of an initial response taking about 90 minutes makes the tradeoff visible. That investment is most defensible for recurring questions with stable answers; immediate response speed can suffer. DNS illustrates the next step. Domain Connect replaces manual record entry with an authorization flow, but provider participation, maintenance and fees limit its reach. Teams adopting that pattern still need a usable path for unsupported providers.

Email evaluation should bring reputation and measurement into the purchase decision. Product checks and automatically generated plain text are candidates for sensible defaults, but Rocha’s claim of improved deliverability is unquantified and should be tested in the buyer’s setting. Open rates require particular care: transparent tracking pixels are affected by privacy measures and pre-reading, so a recorded open cannot simply stand for human attention. Shared tracking domains introduce another dependency because another sender’s abuse can affect their reputation. Separating transactional and marketing streams through subdomains follows Rocha’s practice, while buying a similar-looking replacement domain may introduce phishing-like signals rather than solve the problem.

For content teams, build a set of actual questions before judging an AEO change. Check whether models can retrieve pricing rules, comparison conditions and chart values, then assess answer correctness separately from recommendation frequency. Publishing text also creates an obligation to keep it consistent with the rendered page. A single corrected answer supports a limited conclusion. The host’s CMO-agent workflow suggests how to organize further investigation: connect analytics and code, retain experiment history, propose a hypothesis, obtain human approval, set an experiment window, and monitor execution. The episode supplies that mechanism, without proving any particular marketing intervention effective.

Finally, lower friction in proportion to the cost of failure. Anonymous use, provisioning and claims remained unresolved in Rocha’s account because email abuse can impose costs on other senders. A bounded task with explicit identity or usage conditions is a more assessable experiment than treating registration removal as an automatic requirement. The same discipline applies to buying measurement software. Rocha says four AEO subscriptions still leave him wanting actionable advice, while the host qualifies his assessment by acknowledging limited expertise. Reports can help select a question; choosing an intervention still requires a hypothesis, an acceptable cost and a way to judge the result.

Source

• Original episode: Designing for Agents — Zeno Rocha, Resend