Summary
Fahd Mirza integrates the locally-running Qwen 3.6 35B-A3B model with OpenClaw to test its full agentic capabilities beyond code generation. The task is intentionally designed to exercise the complete agentic loop: given a single natural-language prompt, the model must create project directories, run shell commands, execute npm install, start a live Vite dev server, manage Git, and debug errors — all autonomously on a real Ubuntu system with no human intervention between steps.
The target application is a React dashboard for monitoring a cement plant, displaying live metrics like temperature and RPM with color-coded status alerts. After configuring OpenClaw to point to the locally-served vLLM instance (no API key required), Mirza pastes a detailed prompt and lets the agent run. An initial rendering failure due to broken CSS triggers an interesting sequence: OpenClaw, directed by Qwen, automatically clears the browser cache, restarts the browser, reloads the application, identifies the remaining style issues, and applies fixes — all without user input.
The video offers a candid look at the power and the risks of fully autonomous local AI: the agent has complete control over the system, including browser process management. Mirza explicitly flags this as a security consideration and points viewers to previous OpenClaw security videos on the channel. The demonstration is a practical, real-world test of the SWE-bench-style capabilities seen in Qwen 3.6’s published benchmarks, executed entirely on free, locally-hosted open-weight infrastructure.
📺 Source: Fahd Mirza · Published April 17, 2026
🏷️ Format: Hands On Build
