Qwen and local-agent-builder: build a research agent

This AMD-sponsored tutorial builds a local AI research agent with the Qwen model family and local-agent-builder. It explains the design process rather than installing a finished application. The hardware covered includes AMD Ryzen AI Max Series (Strix Halo) and Radeon AI PRO R9700 GPUs.

The speaker argues that a flat agent accumulates page text and tool outputs that slow inference and can weaken answers. The proposed three-tier design separates an orchestrator, searchers and page analyzers. Searchers save pages to disk; analyzers use grep and selective reads to extract relevant passages without loading every document in full. Tool quotas combine limits enforced by the harness with budgets stated in prompts, with a tradeoff between research depth and runtime.

Pi uses the scaffold to generate the application through a VS Code extension, with llama.cpp serving the model. The walkthrough covers Python environment setup, prompt review and concurrency settings. An initial report omits source URLs; a follow-up coding request fixes citation handling. Web research still requires network access despite local inference.

For evaluation, the speaker runs 30 queries three times each with an LLM judge and reports a 79.1% aggregate score. Some complex queries fail or omit details, and one question comes from a dataset the speaker has not validated. The result describes this benchmark, not general research accuracy.