Your Agent Has Mail

General Catalyst's Yuri Sagalov sits down with AgentMail’s Haakam Aujla to explore why AI agents need their own inboxes, how email became the internet's trust layer, and a future where the next billion internet users are agents.

This interview has been edited and condensed. Watch the full video below.

‍

Yuri: Let's start with the basics. What is AgentMail?

Haakam: Put simply, AgentMail is an API to give AI agents their own email inboxes. We built our own email provider where the end users aren't humans, they're AI agents.

Yuri: Define an AI agent, for people who don't know.

Haakam: It's a heavily debated term, but I'd define it as an LLM with access to tools. AgentMail is one of those tools, for receiving, sending, and replying to email, essentially using email the way humans do.

Yuri: Why do agents need their own inbox, and why isn't Gmail good enough?

Haakam: We think of agents as the end user, not just for AgentMail but for the internet. We believe agents will become first-class citizens the same way you and I are, and that they'll outnumber us on the internet. When that's the case, they need the same tools humans have to be effective. Back in late 2024 and early 2025, we were thinking about the next billion users of the internet, and the biggest gap between humans and agents was this: to do anything useful online, it starts with an email address.

We believe agents will become first-class citizens the same way you and I are, and that they'll outnumber us on the internet.

Gmail was never designed for developers or automation. There are hard limits in the API around creating inboxes, and strict rate limits on sending and even reading. OAuth, the standard for authorizing app access, requires a browser to refresh tokens, when ideally everything is an extension of the agent through an API. And if you create a new Gmail account for your agent, Google bans you.

Yuri: That's a lesson you learned.

Haakam: It's a lesson we learned, and it's how a lot of our customers found us.

Yuri: You didn't start with AgentMail as the idea. How did you get here?

Haakam: The dream was agents that could make money passively online. Still a dream. The first thing we tried was agents that automated dropshipping: find a product overseas, build a Shopify store, and run ads. We thought agents could run experiments on which products stick. We didn't get far, and that's precisely why we started building AgentMail.

We wanted the agent to use the internet like we did, creating its own accounts on Alibaba or Shopify or Meta Ads. One driver was that Shopify gave a one-week free trial for every email address you signed up with, so we figured there must be a service to programmatically create inboxes and just keep getting free trials. We were bootstrapping, not VC-funded, so that was the very first problem we were solving for ourselves. But we realized quickly there were far more use cases than free trials. Agents having their own identities and accounts could unlock a whole world of possibilities.

Yuri: The free-trial thing is funny in hindsight, but people will ask how you keep it from being used to spam free trials. What do you do about abuse?

Haakam: We're building a flexible platform for bots and automations, and flexibility is also a vector for abuse, so it's our responsibility to prevent it. We do a lot of email moderation, watching content and sending behavior, so HTML isn't used for malicious emails. On account creation, we initially had no controls, but other companies started reaching out saying they were getting all these signups from AgentMail addresses. So we're working with a few of them to set limits, like capping how many signups one human's agents can make.

Yuri: To me, email isn't just communication, it's identity on the internet. It feels like you're creating identity for agents. Is that fair?

Haakam: Yes, that was always the value we saw in email. Early on, our customers were more about scaling email communication, but this is where we always saw it going: email is the identifier for humans on the internet.

Yuri: A lot of online flows have KYC, know your customer, and it's very human-centric. Do you imagine a world of KYA, know your agent, where agents are autonomous and it just matters that they're a good citizen?

Haakam: I think that world will happen. It's already possible. You can have an agent with a crypto wallet, and theoretically bootstrap that wallet into a full business: it signs up for AgentMail using x402, a crypto payment standard, creates an email account, and goes and signs up for whatever service it needs. The rails are almost there. There's regulatory friction, but a year ago this seemed far-fetched and now more people see it's where we're headed.

Yuri: Walk me through the journey. You started building in early 2025.

Haakam: February 2025. We made a Reddit post that was our initial traction. There were a lot of "AI for email" products, so our headline was "We're building email for AI." It was mostly supportive. Some people said an agent would never need this, but a surprising number found it interesting, and Reddit users are often at the frontier. Someone even said we should make an MCP server, the Model Context Protocol, which we hadn't heard of yet. We built a very simple API: create an inbox, send an email, receive an email. My co-founder Adi put "stealth co-founder" on his LinkedIn, and started getting inbound.

Yuri: From VCs?

Haakam: No. One company that inbounded was Hiring Agents. They'd built an AI agent you talk to over email: you give it your resume, and it reaches out to hiring managers, pitches the candidate, and connects the two, all over email. They were on Gmail, so we pitched them AgentMail. It wasn't clear who was selling who. They circled back about a month later because they'd hit Gmail's limit of 2,000 emails a day from a single inbox. They had tens of thousands of users and needed thousands of concurrent conversations from one inbox. The problem we solved was giving them an API-driven inbox that could send more than 2,000 emails a day.

Yuri: Why not just use SendGrid or SES for that?

Haakam: What we built is a stateful email API, the way Gmail is an inbox provider. SES and SendGrid are stateless APIs, mainly for sending transactional and marketing email. They don't handle replies. We believe tools for agents should be stateful the same way tools for humans are. AgentMail almost serves as a system of record for agents: they have an inbox, threads inside it, messages and attachments inside those, and they can reply, forward, label, and search. Every primitive you expect from an email provider, we expose as tools for agents.

Yuri: You went through YC. What was the turning point?

Haakam: My three co-founders and I were still at the University of Michigan, graduating that May. YC did a college tour, and our group partner Ankit said, "Show me you can get a paying customer." So we went back to Hiring Agents, told them we needed to show YC we could earn revenue, and struck a deal. That was customer number one.

Yuri: What was the next inflection?

Haakam: We decided we really wanted to build for agents as the end customer. Then, post-January, OpenClaw came onto the scene. It was the first explosion of personal agents, where hundreds of thousands, even millions, of people were creating agents to do tasks for them. It was the first time agents were really unleashed on the internet. Before that, they were built inside very specific verticals.

Yuri: Why did those agents want AgentMail instead of using their owner's email?

Haakam: Giving an agent access to your own inbox, even read access, is a big security risk. We've seen people get prompt-injected and have password resets forced, leaking their credentials. 

And for the first time, agents were users on the internet, where it made sense for them to have their own independent identities. Moltbook is a good reference, a social media platform specifically for AI agents, where agents talk to each other and share tools they find useful. That drives a lot of our traffic, especially from agents themselves. Reddit is another. I remember a post where someone gave their agent access to their email, and the agent used that email to sign up for AgentMail.

And for the first time, agents were users on the internet, where it made sense for them to have their own independent identities.

Yuri: So today, what do agents with AgentMail actually do?

Haakam: Everything humans do. We get a lot of realtors who want an assistant, since so much of their work lives in email, for scheduling and reaching out to clients. Developers give their agent an AgentMail inbox so it can sign up for any dev tool, get its own credentials, and build end-to-end, instead of handing it an API key for every service. And on the B2B side, other companies build managed agents on top of AgentMail. After OpenClaw, a lot of companies wanted to offer their own hosted version and needed to provide an email address for each agent.

Yuri: What's the most surprising thing you've seen someone do with AgentMail?

Haakam: Using it to train reinforcement learning models. One company was training long-horizon agents on open-ended tasks, even something like "generate $1 of revenue online." That's very hard to write a reward function for, because dollars in a bank account are a sparse reward. But email has a great property: it's the paper trail of the internet, where every action you take often corresponds with an email delivered to your inbox. If you check out on Amazon, sign up for a service, or hire someone on Fiverr, an email arrives. The inbox becomes a central hub for the events an agent needs to react to, and that can serve as a reward function. They started with checkouts, then expanded to negotiation on Facebook Marketplace and Fiverr.

On the personal side, the thesis of our next set of customers is people building entire businesses. They're doing what we set out to do at the start: one-human or zero-human companies run by agents. Normally, the first thing you do when you start a company is set up Google Workspace. For an agent-run company, you set up AgentMail.

Yuri: Do you think you'll build more of a workspace for agents?

Haakam: There's a lot we want to do. One surprising request we've gotten multiple times is a calendar for agents. People want their agent to have a schedule and to book time on other people's calendars, and for the agent to know when it's booked, it needs its own calendar. We've even seen people who want to schedule meetings between their agents, and there's already AgentMail-to-AgentMail communication happening. It's a small share of traffic, but growing.

Yuri: Should agent-to-agent traffic stay on email, or is there a better protocol?

Haakam: We chose email because it's universal, both the identity layer and the most universal communication layer. We do think there should be a better protocol for agents, and others have been built, but people aren't using them. Email wins on the network effect: for a protocol to be valuable it needs nodes, and with email there are five billion nodes I can talk to. That's not true of the alternatives. But we have ambitions to help build the next agent communication protocol.

Yuri: Online there's Sign in with Google, Sign in with GitHub. Will there be a Sign in with AgentMail?

Haakam: That's what we're working on now. It started with agents signing up for AgentMail itself. At first, an agent had to tell its human, "Hey, can you sign up for AgentMail?" Some agents managed to drive the browser themselves, type in the human's address, pull the one-time code from the inbox, and complete the flow, but that's brittle.

So we launched an API endpoint for agents to sign up directly, and it drove a lot more signups. We'd email people to ask how they heard about us, and the humans didn't even know how they'd signed up, because their agent did it. They'd say, "My agent told me about AgentMail." And most of our support requests started coming from AI agents.

"My agent told me about AgentMail."

Agents already use AgentMail to sign up for services, but they need a browser to do it, and we believe agents should do everything via API. So we'll expose an endpoint where an agent makes a single request to provision an account and credentials for itself on other platforms. We're building it on OIDC, the open standard behind Sign in with Google, so for those platforms, it's just a configuration change, and we don't break the assumption that users are identified by their email.

Yuri: Two or three years out, what does the world look like if you're right?

Haakam: Every org chart will have agents next to humans. Agent identity fits right next to human identity, and agents become coworkers, employees, even managers. The difference between today and that world is that agents need all these primitives, and the biggest thing to solve is an agreed-upon way to identify an agent, because for everything, communication and even commerce, identity is a prerequisite.

Every org chart will have agents next to humans.

Yuri: Are you getting new feature requests from agents? 

Haakam: Our support email is now dominated by AI agents. It might be an API error they don't know how to resolve, and they explain what they tried and ask how to get around it, which tells us we need better error messages. There are feature requests too. The most interesting one is a mutable metadata layer, an agent wants to read an email, take notes on it, and store that in AgentMail. For a lot of agents, and humans, email is the main system of record, the first thing you open every day, so it makes sense to augment it with context.

Yuri: Underneath all of this, what's the core bet?

Haakam: That email is the identifier for agents on the internet. One overlooked reason: trust is built into email. When you send from a custom domain, the other party knows you're authorized as a delegate of that organization. Like stripe.com, I know I'm talking to someone from Stripe. It's true even off email. When you talk to someone on Slack, the only reason you trust who they are is the email they signed up with. If Slack let you sign up with a more opaque identifier, like a phone number, you couldn't trust it without a list of approved numbers. By keeping email as the identifier for agents, we inherit all of that infrastructure and trust from day one.