When Your AI Agent Runs for 16 Days Straight

Most AI coding agents top out at short, interactive sessions - a few minutes of back-and-forth before context degrades. At Factory, we built a system that sustains autonomous software development for hours, days, and even weeks. An orchestrator decomposes large projects into features, spawns focused workers with clean context, coordinates through git, and validates at every milestone. This talk covers the architecture behind multi-day agentic workloads: managing context limits, multi-model orchestration, balancing serial reliability with parallel speed. We also share production data on what actually changes when agents shift from 8-minute conversations to 16-day autonomous runs.