essay

A self-driving company starts before the ticket

Jeff Reynar
A self-driving company starts before the ticket

Amjad Masad, Replit’s CEO, told the audience at TechCrunch Disrupt on October 16 that Replit is becoming a self-driving company. “The company is sort of self-improving,” he said, “so a lot of the work that’s going on is driven by agents.”

How Replit runs

His example was support. A ticket goes to an AI agent first, which triages it against GitHub issues, the codebase and Linear. People review what the agent found and decide what to do, and when they approve a fix, agents deploy it with a click. “You need human judgment; you need taste,” Masad said. In his telling, people spend more of their time on architecture, strategy and deciding priorities.

That’s the right design: agents do the digging, and people make the call. It’s close to what Y Combinator described this year as the self-improving company.

Where the loop starts

Replit’s loop starts at the ticket, but the problems a company needs to fix don’t all arrive as tickets. A customer emails their account manager that an export keeps failing. A partner mentions in a shared Slack channel that the API started returning errors overnight. A prospect tells a salesperson that the integration they need doesn’t work. Each report reaches someone, and it becomes a ticket only if that person stops to file it.

Imagine the account manager is in meetings all afternoon and files the export problem two days later. The agent that triages tickets starts two days late. It’s fast, but it only sees what reaches the ticket queue.

So a self-driving company needs something at the front that reads where reports actually arrive: the inboxes and channels of the people customers talk to.

How to automate it

AI reads the company’s messages on its behalf, notices that the customer’s email describes a failing export, and turns it into a task that support and engineering can both see, with the email attached. A workflow checks it against open issues and, if it’s new, files it and assigns it to whoever owns exports. From there the loop looks like Replit’s: an agent investigates, a person approves the fix, and the customer hears back.

That’s what we’re building at this+that. DoBox reads the messages a company already gets across Gmail, Outlook, Slack, Microsoft Teams, and Google Chat, and turns them into tasks visible to everyone who needs to see them, then hands each to a workflow that handles it or to the person who should handle it. From a task, you can delegate the work to an AI agent, Claude Code or Codex on your own Mac or our cloud agent, and it comes back with a draft pull request for a person to review. Teams assemble the support-to-fix pipeline from workflows; we built one for our own support mail as an example.

Try it on last month’s fixes. Pick a few bugs that customers reported, find where each report first arrived, and count how many started as a message to someone rather than a ticket. If the count surprises you, see how delegation works.