Summary
Bart Slodyczka demonstrates a complete end-to-end build of a locally-run customer support agent using the Pi agent harness (pie.dev), Qwen 3.6 35B via LM Studio, and Zendesk as the target inbox system. Pi is presented as a lightweight alternative to more comprehensive systems like Open Claw or Hermes Agent — suitable when you need a focused, regulated task agent rather than a general-purpose AI with full computer access.
The build starts from a blank VS Code folder: installing Pi, connecting it to LM Studio via the Pi LM Studio extension (which surfaces all locally-running models including those across multiple machines via LM Link), and then using conversational prompting to have the Qwen 3.6 35B model scaffold the entire agent package — including a Zendesk extension file, tools definitions, prompt templates, and a response quality skill — entirely from its own documentation. Slodyczka walks through retrieving a Zendesk API token, setting up environment variables, and reviewing the auto-generated folder structure.
The video is candid about real-world friction: the model entered a loop during environment file creation, requiring a manual interrupt — a known issue Slodyczka attributes to smaller models. The finished agent is structured as a portable package, meaning it can be moved from a local machine to cloud infrastructure or a dedicated device for production deployment without modification. The full setup runs on a Mac Studio with no cloud API calls required.
📺 Source: Bart Slodyczka · Published July 06, 2026
🏷️ Format: Hands On Build
