Type: url
Original: https://addyosmani.com/blog/long-running-agents/
Status: generated
Summary
This Addy Osmani post surveys “long-running agents” — AI agents that keep making progress across many sessions and sandboxes over hours, days, or weeks — distinguishing long-horizon reasoning, long-running execution, and persistent agency, and naming three recurring obstacles: finite context, no persistent state, and no self-verification. It compares how Anthropic (harnesses, then a brain/hands/session split), Cursor (planner/worker/judge roles), and Google (the Gemini Enterprise Agent Platform, including Agent Runtime, Agent Sessions, and Agent Memory Bank) each address those problems, and sets out five production patterns — checkpoint-and-resume, delegated approval, memory-layered context, ambient processing, and fleet orchestration — along with practical guidance and current limitations (cost, security, alignment drift, verification, and the human role).