Simon Last (Loom co-founder) documented a method for running AI coding agents continuously for 13 days on real production work. The approach treats the agent as an unreliable worker and builds disciplined infrastructure around it. Key principles include decomposing work into atomic file-scoped task files with explicit acceptance criteria, using CLAUDE.md as persistent project memory committed to the repo, writing a bash orchestration script that feeds tasks sequentially to Claude Code CLI in isolated sessions, gating every task with automated tests before committing, and committing after every completed task for easy rollback. The article provides a full orchestration script, task file templates, CLAUDE.md structure, failure mode diagnosis, and advanced techniques like parallel agent tracks and self-updating memory files.
Table of contents
How to Run AI Coding Agents Continuously for DaysTable of ContentsWhy AI Coding Agents Fall Apart After an HourWhy AI Coding Agents Lose the PlotThe Core Architecture for Multi-Day Agent RunsImplementing the Task Loop SystemComparison: Typical Agent Limits vs. Extended Runs with Survival TechniquesAdvanced Techniques for ResilienceCommon Failure Modes and How to Diagnose ThemCommon PitfallsKey Takeaways: Building Systems, Not PromptsSort: