Why Claude Keeps Recommending Resend: What Agent Experience Design Means for Every Developer Tool
Claude Code recommended Resend first in 64 of 102 email API suggestions in a research study, raising a critical question about how AI agents pick vendors. Latent Space's interview with Resend CEO Zeno Rocha unpacks the design decisions — from machine-readable pricing pages to MCP tool architecture — that shape what Claude recommends and why it matters for Indian teams building on AI-assisted workflows.
When Claude Picks Your Vendor Before You Do
Something quietly remarkable happened in a research study that Latent Space’s discussion of this topic covers in detail: when developers asked AI coding agents to implement email sending in their applications, Claude Code recommended Resend first in 64 out of 102 email-related recommendations the researchers could extract. SendGrid, a much older and larger incumbent, appeared first only seven times — even though it showed up 55 times as an alternative. Zeno Rocha, Resend’s co-founder and CEO, joined the Latent Space podcast (https://latent.space/p/resend/) to explain both sides of that result.
The question this raises is not just about email APIs. It is about a fundamental shift in how software tools get discovered, evaluated, and adopted — and Claude is sitting right at the centre of that shift.
What Is ‘Agent Experience’ and Why Should You Care?
You are probably familiar with SEO — shaping your content so that Google’s algorithm surfaces it to human searchers. Answer Engine Optimisation, or AEO, is the newer practice of shaping your content so that AI models surface it when someone asks a question. But according to Latent Space’s discussion of this topic, Zeno Rocha has been thinking beyond AEO since early 2025, coining the term ‘Agent Experience’ to describe what happens after the AI has picked your product.
Agent experience is about designing your product — its documentation, its pricing page, its APIs, its tooling — so that an AI agent can not only discover it but actually use it successfully on a human’s behalf. This is a subtle but important distinction. AEO gets you recommended. Agent experience determines whether the agent can complete the task without breaking down.
If you are a product manager, a startup founder, or someone who oversees technology decisions at an Indian company, this matters to you even if you never write a single line of code. The agents your developers are increasingly relying on — Claude Code being the most prominent example — are making vendor choices based on how well a product communicates with machines, not just with humans.
The Pricing Page That a Machine Can Actually Read
Here is a concrete example from the Latent Space interview that illustrates what agent-friendly design looks like in practice.
Resend’s pricing page has a slider. A human visitor can drag it to their expected email volume and read off the cost. That is a perfectly good user experience for a person. But a slider is essentially invisible to an AI agent trying to compare costs across providers. So Resend made a small but significant change: add .md to the pricing URL and the agent gets a plain-text table it can parse in a single request.
That one decision — exposing a machine-readable version of the pricing page — means that when Claude is asked to compare email providers for a developer, it can actually answer the question accurately for Resend. For a competitor whose pricing lives inside a JavaScript-rendered component or a drag interface, Claude may simply be unable to retrieve comparable data.
Now consider what this means for an Indian SaaS startup based in Pune that has built a B2B invoicing tool. If the team uses Claude Code to accelerate development and an engineer asks the agent to ‘add email notifications for payment reminders,’ the agent will reach for whichever email provider it can most confidently evaluate and implement. If your chosen provider’s documentation is thin, inconsistently structured, or machine-hostile, the agent will route around it — not out of bias, but out of practical necessity.
The 80-Tool Problem in MCP Design
Resend also runs an MCP server — a Model Context Protocol integration that allows Claude and other AI agents to interact with Resend’s platform directly through tool calls. According to the Latent Space discussion, Resend’s MCP server had grown from roughly 20 tools at launch to more than 80 by the time the interview was recorded. The goal, as Zeno describes it, is for someone working in a chat interface to be able to do everything they can do in the Resend dashboard: send an email, invite a teammate, check usage, manage billing.
This is an admirable goal, but the interview surfaces an honest design tension. Zeno describes opening the collapsed tool call logs in his own agent sessions and finding that some tool responses were so large they filled the available context window. That is a real problem: when a tool returns more data than Claude can hold in memory at once, the agent’s ability to reason about it degrades. Adding more tools does not fix a bloated response from a single tool. These are separate problems.
Anthropics’s own guidance on writing tools for agents — cited in the Latent Space show notes — addresses exactly this: returning less data when a tool result risks filling the context window. The lesson for anyone building integrations with Claude is to think carefully not just about what your tool can return, but about what the model actually needs from each call.
Compare this to Cloudflare’s approach, which the Latent Space discussion contrasts directly with Resend’s. Cloudflare’s Code Mode MCP server exposes its API through just two tools: search() and execute(). That is a radically different philosophy — fewer, more flexible tools rather than a large catalogue of specialised ones. Neither approach is universally correct, but as a non-technical decision-maker, you should know that the design of an MCP server shapes how capable and reliable Claude will be when working with that platform.
A Scenario Closer to Home
Imagine a growth team at a fintech startup in Bengaluru. They use Claude to help write and schedule transactional email campaigns — OTP messages, loan approval notifications, KYC status updates. Their developer, stretched thin across three products, asks Claude Code to help integrate an email sending service. Claude suggests Resend.
The developer accepts the suggestion, the code gets written, and within the hour the integration is live. At no point did the team run a formal vendor evaluation. The AI agent made a recommendation, the developer trusted it, and that became the default. This is not a hypothetical edge case — it is increasingly the standard workflow in developer-led teams across India’s startup ecosystem.
This is why Zeno Rocha’s framing matters. The question is no longer only ‘how do we win a human’s attention.’ It is ‘how do we win the agent’s confidence.’ And the agent’s confidence is built through documentation quality, structured data availability, reliable tooling, and predictable API behaviour — not through sales calls or banner ads.
What This Cannot Do Yet — Honest Limitations
The Latent Space discussion does not shy away from the edges of this picture. Email deliverability remains a genuinely hard infrastructure problem that no amount of agent-friendly design fully solves. Resend introduced custom tracking domains partly because shared tracking domains — where multiple customers’ links route through the same domain — create shared risk. If one sender on that domain gets flagged for phishing, it can affect everyone else’s click-tracking links. That is a real operational concern for any Indian company sending transactional email at scale.
The AEO study that showed Claude Code preferring Resend also showed a more competitive picture in a later, more comprehensive study, where Postmark came out on top for most models. Model preferences are not fixed. They shift as training data changes, as documentation improves for other providers, and as new model versions are released. A product that benefits from Claude’s current preferences has no guarantee those preferences persist.
And practically speaking, MCP integrations — including Resend’s — are still evolving. The context-window problem with large tool responses is not yet solved. The 80-tool catalogue means more surface area for things to behave unexpectedly in long agent sessions.
What to Watch Next
According to the Latent Space interview, Resend had 4.3 million users and 120,000 paying customers at the time of recording. That scale suggests the agent-experience approach is producing real adoption results, not just theoretical benefits.
If you manage a product, a developer team, or a technology budget in India, the practical next step is straightforward: open your own product’s documentation or pricing page and ask Claude a customer question about it. See whether Claude can answer accurately. If it cannot, you have found your first agent experience gap — and it costs nothing to fix a documentation page.
