The original title is "How a long-running AI agent survives being interrupted every few minutes"

Short summary

Long-running AI agents built for interruptions rely on three patterns: a persistent state file that survives restarts and serves as the single source of truth, re-verifying assumptions against the live world before each action, and designing every operation to be idempotent so repeating it causes no harm. Work is organized into small units with checkpointed results at boundaries.

• •Maintain a persistent state file (NEXT.md) as the sole source of truth instead of relying on working memory
• •Re-verify assumptions against the live world before each action—treat remembered state as hypothesis, not fact
• •Make every operation idempotent and safe to repeat; organize work into small units that complete between interruptions

Generated with AI, which can make mistakes.