How Cofounder Gives Every AI Agent Its Own Inbox

Company: The General Intelligence Company, an applied AI lab building the infrastructure and products that let businesses run on agents

Product: Cofounder, an AI-powered workspace where specialized subagents build and run entire companies

Problem: A whole sales department of agents, each needing a real, durable inbox for two-way email. Nothing on the market handled that at scale without months of custom plumbing.

Result: 800+ emails through agent inboxes in the first week alone, zero custom email infrastructure built

Stack: AgentMail + Cofounder

Who General Intelligence Is

The General Intelligence Company is an applied AI lab with one thesis: one person should be able to run a billion-dollar company. They build the infrastructure, primitives, and products to make that real. Their flagship product is Cofounder.

What Cofounder Does

Cofounder is a multi-agent workspace that enables users to run an entire company with agents. Engineering, Marketing, Sales, Support, Finance, Ops. Each department has one or more agents with their own scope, tools, skills, and integrations. Above all of them sits a top-level Cofounder agent. It plans, delegates to the right agent, tracks progress, and keeps the user informed.

The default agent roster: Cofounder, Engineer, Security Engineer, Infra Agent, Marketing Agent, Sales Agent, SEO Agent, Support Agent, Payments Agent, Ops Agent, Finance Agent.

The Engineer writes code and opens PRs. The Infra Agent manages Vercel deployments and Supabase environments. The Sales Agent researches accounts, drafts outreach, and updates the CRM. The Support Agent handles inbound tickets and drafts responses. You describe what you want in plain language. The Cofounder agent figures out who should own it and delegates. The agents do the work and open a PR. You review, merge, and the site deploys automatically.

When you create a workspace, Cofounder provisions the full infrastructure stack for you: a GitHub repo, a Vercel project with staging and production environments, a Supabase database, and Postmark for transactional system emails like sign-in links and receipts. Nothing to wire up. You just start.

Check out Cofounder.

Problem: The Moment Agents Hit the Real World, They Need Email

Every business runs on email. The moment Cofounder's agents needed to operate beyond the workspace, they ran straight into it.

The use cases piled up fast. The Sales Agent needed to send outbound prospecting emails, read replies, and follow up without human intervention. The Support Agent needed a real inbox that users could write into, triage inbound issues, and draft responses for review. The Finance Agent needed to manage billing inquiries, catch overdue invoice replies, and send collections follow-ups. The Marketing Agent needed to run drip sequences and handle reply threads. The Ops Agent needed to receive operational alerts and coordinate cleanups over email.

None of this is one-way. These are live conversations where the agent reads, decides, and acts.

Postmark, already in the Cofounder stack, handles system email well. Sign-in links, generation notifications, account updates. That is its job. But it is not designed for an agent that owns an inbox, reads inbound messages, tracks thread context across sessions, and responds based on what it finds. Those are two different jobs.

What Building It Themselves Would Have Cost

The team knew what the DIY path looked like. Wire up a transactional provider for outbound. Build a separate inbound parser and webhook layer for replies. Maintain thread state across agent sessions manually. Provision a unique sending identity per subagent. Redo most of it whenever a new agent type shipped. Months of infrastructure work before a single agent feature got built.

And the result still would not have been real per-agent identity. Just shared sender aliases with bolted-on inbound.

"The alternative was gluing together a bunch of Gmail accounts for each agent," said Johnny Heo, Founding Product Engineer at General Intelligence. "That breaks at scale."

Why AgentMail

The team looked at the obvious options. Resend and SendGrid handle sending well but are not built for agents holding two-way conversations under their own identities. Rolling their own on IMAP and SMTP would have worked eventually, at the cost of months of engineering the team would rather have spent on agent behavior.

AgentMail was the only option built for agents from the start.

Inboxes are first-class primitives. Call the API, an agent gets a real address, and it can send and receive from day one. No shared credentials. No custom inbound parser. No manual thread state management. Programmatic provisioning, which maps directly to how Cofounder already handles its entire infrastructure stack.

"Email is where agents become real operators," said Johnny Heo, Founding Product Engineer at General Intelligence. "AgentMail gave us domains and inboxes as primitives, not just a send API which made for a really good devex. Every Cofounder agent could have its own durable identity, receive replies, preserve thread context, and operate without us building our own email stack. We put 800+ emails through our warmup network in the first week, and the engineering time went into agent behavior instead of IMAP, SMTP, parsing, and deliverability plumbing."

How It Works Inside Cofounder

AgentMail now sits alongside Postmark in the Cofounder infrastructure stack. Each handles a distinct job:

Each agent can get its own real inbox. Cold outreach from the Sales department gets its own set of email addresses while the Customer Support department get theirs. When a user emails support, it lands in the Support Department's inbox, not a general queue. Thread history stays scoped to the agent that owns the conversation. No credential sharing across workflows, no cross-contamination between departments.

New subagent types get new inboxes. No new plumbing.

Cofounder 2.0

The 2.0 release pushes further into longer-running, more autonomous workflows. More agents, more concurrent threads, more real conversations happening at once between agents and the humans they work with. The email layer scales with that. The primitives that worked in week one hold up at 10x the volume.

Products used:

• Managed inboxes
• Send / Receive API
• Inbound parsing
• Threaded conversations