Pi Durable EXPLAINED: Build Always-On Agents with Local Models

Pi Durable is a building block for always-on agent applications, not a new coding-agent CLI. Learn how checkpoints, conversations, and storage support recovery after interruptions, where Durable fits in the Pi ecosystem, and the six steps to initialize a harness. Then follow a local llama.cpp demo with Qwen3.5-0.8B, including a second script that resumes interrupted work from SQLite. The APIs are experimental and may change; tool replay safety and restarting the host process remain important application responsibilities. --- 🔗 *Links* - Tutorial companion code (fresh chat and restart recovery): https://github.com/alejandro-ao/pi-durable-demo - Pi Durable introduction: https://earendil.com/posts/pi-durable/ - Pi source code: https://github.com/earendil-works/pi - Pi: https://pi.dev/ - llama.cpp: https://github.com/ggml-org/llama.cpp - Qwen3.5-0.8B: https://huggingface.co/Qwen/Qwen3.5-0.8B - GGUF weights used in the demo (BF16): https://huggingface.co/unsloth/Qwen3.5-0.8B-GGUF --- 🤓 *Topics Covered* - Durable agent checkpoints and crash recovery - Harness architecture, conversations, and storage - Local-model setup and restart recovery demo --- ⏱️ *Timestamps* 0:00 Introduction and tutorial roadmap 1:48 What Pi Durable is and crash recovery 8:38 Harness versus application in the Pi ecosystem 13:46 Building an application around Pi Durable 19:30 Six steps to initialize the harness 23:11 Local-model code demo, restart recovery, and outro 👋 *Connect with me* My website: https://alejandro-ao.com/ X (Twitter): https://x.com/_alejandroao LinkedIn: https://www.linkedin.com/in/alejandro-ao-