Day 1 – April 22, 2026
Long-Running Agents: The AI that never sleeps
Today's AI agents are impressive but often don't run for long - they forget context mid-task, hallucinate when sessions get long, and declare victory before the job is done. What if an agent could work autonomously for hours or even days, maintaining state, recovering from failures, and actually shipping results?
In this session, we'll explore the architecture and patterns behind long-running agents — persistent AI workers that go far beyond single-prompt interactions. Through two live demos built on Google ADK and Cloud Run, you'll see a 24/7 competitive intelligence agent that monitors competitors across Reddit, YouTube, Hacker News, and the web (with historical trend detection powered by Firestore), and a creative coding agent that transformed a large Blender scene into a browser-explorable 3D environment in a single continuous session.
We'll cover the key engineering breakthroughs that make this possible — agent harnesses, persistent memory patterns, self-verification loops, and how Agent Engine's Sessions and Memory Bank provide managed infrastructure for agent persistence at scale. You'll leave with practical patterns for building agents that don't just answer questions, but plan, execute, debug, iterate, and ship.
Read more