Mar 21, 2025 · 44m · a16z
Automating Developer Email with MCP and AI Agents
0:00 / 0:00
▶ Watch on YouTube →
gold bands on the timeline = statements, start to end. Hover to read, click to jump. CC turns on captions
Zeno Rocha, Founder and CEO of Resend, joins a16z Partner Yoko Li to discuss the shift from Developer Experience (DX) to Agent Experience (AX). They explore how email infrastructure, Model Context Protocol (MCP), and software design principles must adapt to empower autonomous AI agents in modern software workflows.
How this conversation actually went
Every chapter scored 0–10 on four independent dynamics. Hover any point for the reasoning behind the score. How this is scored →
The host as informed peer 4.4 Guest teaching 3.6 Guest disagreement 0.0 The host pushing back 0.0
speaking balance: gold is the host, purple is the guest (3 minute bins)
Sharpest disagreement ▶ 18:31 Zeno questioning agent classification boundaries
In a very gentle counter-point, Zeno hesitates when asked if the generation workflow is truly an agent, noting that in some ways it is but in other ways it might just be a focused tool script.
Hardest push from the host ▶ 24:14 Yoko questioning MCP local-first limitationsYoko pushes past superficial enthusiasm for MCP to highlight architectural pain points like Server-Sent Events (SSE) and asks what is missing to make MCP operate beyond local command executions in true production ecosystems.
Biggest teaching moment ▶ 34:23 Stripe's advice on MCP design strategyZeno educates the audience on proper API-to-MCP design, relaying Stripe's insight that teams should not blindly dump OpenAPI specs into MCP servers but rather design specifically around high-value user workflows.
The host holds their own ▶ 19:34 Yoko articulating the technical definition of an agentWhen Zeno asks for her definition, Yoko shows crisp technical authority by breaking down an agent into a multi-step LLM execution process operating like a system-level process with model decision-making in the loop.
the scores for every segment, with the reasoning behind each
| Chapter | Topic | The host as informed peer | Guest teaching | Guest disagreement | The host pushing back | Why |
|---|---|---|---|---|---|---|
| Transitioning from Developer Experience to Agent Experience | 3 | 3 | 0 | 0 | Host Yoko Li sets up the overarching topic of transitioning from DX to Agent Experience (AX). Zeno explains how products like Resend must remove friction such as manual verification delays to cater to automated agent consumers. | |
| Adapting Email Interfaces and Protocols for AI Senders and Receivers | 4 | 4 | 0 | 0 | Yoko and Zeno discuss how email protocols adapt to AI agents. Yoko cites LLMs.txt as a parallel format optimization, while Zeno highlights why plain text emails save tokens and parse faster for agent-to-agent communication. | |
| The Text-to-App Paradigm and Email Customization | 4 | 3 | 0 | 0 | Yoko references her daily use of tools like Cursor to prompt a discussion on text-to-app paradigms. Zeno shares how user signups shifted to AI-first builders like Lovable and Bolt, pointing out that publishing is the primary call to action. | |
| Modernizing Email Creation with React Email and New Email | 5 | 4 | 0 | 0 | Yoko illustrates creative iteration using an oil painting blank-canvas analogy and LoRA style presets. Zeno explains React Email and New Email, noting how system prompts can inadvertently constrain AI creativity. | |
| Defining AI Agents and Multi-Step Execution | 7 | 3 | 0 | 0 | Yoko demonstrates strong technical mastery by defining AI agents as multi-step LLM execution processes and probing the limits of local-first MCP implementations, local CLI versus runtime workflows, and data gravity. | |
| Creative AI Use Cases, Automated Newsletters, and Programmatic Domains | 4 | 4 | 0 | 0 | The pair explore novel agent applications, including automated AI newsletters and programmatic domain creation. Yoko shares a lighthearted personal story about acquiring her surname as a top-level domain (.li). | |
| Advice for Developers and the Dracula Theme Origin Story | 4 | 4 | 0 | 0 | Zeno offers advice for developers converting to AI engineering, passing along insights from Stripe's MCP team. He then recounts the origin story of his popular Dracula color theme created while hospitalized in Madrid. |
Statements from this episode (22)
Rocha: Most Supabase databases and Resend emails were created by humans
“When you go to all these different apps, you go to Supabase, the most databases you see there were built by humans. You go to Resend, the emails were sent by humans or drafted by humans and then sent programmatically.”
Rocha: AI agents will soon perform majority of digital actions over humans
“I think we're going to see a big shift in terms of who is the actor, who is the creator. And I believe it's going to be the majority of the actions will be taken by agents instead of humans.”
Rocha: SaaS companies should reevaluate blocking bot signups via reCAPTCHA
“So even like small things like adding reCAPTCHA on the signup, To prevent bots. Now you have to think, do you really want to prevent bots from signing up? Maybe you don't, right?”
Rocha: Legacy email services require a two-day manual verification process
“You go to SendGrid, you go to Postmark, to Mailgun, you sign up for these services, and the first thing you get is a manual verification process that takes two days for you to get approval of your account, right?”
Rocha: Autonomous AI agents will never tolerate multi-day onboarding delays
“In an agent-first world, that's simply not, you know, possible. Like an agent will never wait two days to take an action.”
Rocha: Documentation investments directly improve LLM output quality
“The more you invest in docs and knowledge base, The more quality of the LLMs that you're going to have, like in terms of the output there.”
Rocha: Plain text email is superior to HTML for AI models
“Turns out in today's world, that's actually better than the HTML version because it's easier to parse. You're using less tokens. It's cheaper.”
Rocha: AI agents will communicate directly with each other via email
“I guess the difference now is that we're going to be living in a world where agents are talking to other agents and these agents are sending emails, but also receiving and parsing and sending to another inbox that is also powered by agents that's parsing, taki…”
Rocha: Every industry will have specialized text-to-app AI tools
“I think we're going to see that across the board for every single industry. There's going to be one of those. The output might be a little bit different but you're always going to need something that is tailored to that industry.”
Rocha: Publishing live is the real aha moment for text-to-app tools
“And for Bolt, Lovable v Zero, it's publishing. It's going live with the website. That's the aha moment. It's not generating the code or building the to-do app. It's actually putting it live.”
Rocha: Basic LinkedIn customization is insufficient for AI SDR tools
“You see all these SDR tools and you know, it's great that we're seeing a revolution there, but just customizing based on your last LinkedIn message, it's not going to be enough.”
Rocha: Uber designer spent two weeks prototyping with Cursor instead of Figma
“This is a designer that works at Uber, and he was telling me how, like, for the past two weeks, he didn't open Figma. He was just building all these prototypes using Cursor, using vZero, using these tools that are now available for folks.”
Rocha: Resend's prompt forced Apple-style email designs, limiting prompt creativity
“And that's because on the system prompt, we told, you know, the LLM to build Apple-like emails with curved borders and black and white vibes versus, you know, go crazy.”
Rocha: Resend uses separate AI agents for template building and sending
“In the case of new email, for example, we have One agent to build the email template and then there's another one to actually send it or schedule it.”
Li: An AI agent is a multi-step execution process centered around an LLM
“I think of it as a multi-step LLM execution process. It's almost like a process in the systems level. The difference is that you have LLM in the middle to make the decisions.”
Rocha: OpenAI adopting MCP would make it the de facto AI protocol
“Imagine if open AI adopts that, then it's over, you know, like that will be the de facto protocol.”
Rocha: System-of-record apps will dominate AI agent workflows
“I think all of the system of records type of applications, they will be front and center at this, you know, new revolution, like, because all of your issues are already on linear.”
Rocha: Text-to-app tools will compete on database setup speed
“And I think the race with those is going to be around, like, who can reduce friction the most? Who can spin up a new Postgres database faster?”
Rocha: Build MCP servers around core use cases, not full API spec
“Don't start through, like via the API, like just grabbing your open API spec and generating an MCP server. No, start from the use case. What are people actually doing with your product? And then you build, like, you don't need full coverage of your API, just t…”