Why Your Overnight AI Agent Fails (And How Episodic Execution Fixes It)

Why Your Overnight AI Agent Fails (And How Episodic Execution Fixes It) How I built Nightcrawler -- an autonomous agent loop that runs Claude Code for 12+ hours with structured handoffs, crash recovery, and 8 termination conditions. You go to bed. Claude Code is humming along, refactoring your test suite. You set it loose with --dangerously-skip-permissions and a prayer. You wake up 8 hours later. What do you find? If you are lucky: a completed task. If you are unlucky -- and I was unlucky many times -- you find one of these six disasters. The 6 Death Spirals of Long-Running AI Agents Every developer who has tried to run an AI coding agent overnight has hit at least one of these: 1. The Context Cliff After 30-60 minutes of work, the model's effective context fills up. It starts forgetting what it did earlier. It re-reads files it already processed. It contradicts decisions it made 20 minutes ago. Eventually it starts undoing its own work. 2. The Hallucinated Handoff The agent writes "I