The most technically interesting part of this was building a self-improving AI loop Traditionally we think of coding agents as tools to do work faster, but we are quickly moving into a world where async agentic work is fully decoupled from human work This means you can build an AI system that is fully self-improving & can perform experiments and react to results. Based on OpenClaw-style heartbeats, we built a similar system using a bunch of Devin automations: - Calling: Agent dialed restaurants in a morning and an afternoon window on weekdays. - Review: Agent read every transcript afterward and edited the phone agent's prompt. - Idea: Agent once a day, picked one new thing to try. - Webhook: When money or a call came in, an event woke a fifth kind of Devin to handle it. Every ~20 minutes a heartbeat ran in the form of a Devin automation. Each heartbeat is its own Devin session + its own sandbox VM. They all begin by bulling down the "brain" from s3 Read more about the architecture in the quoted article