Open World News

This week's featured content highlights two distinct approaches to advancing AI systems, from open-source agent architecture to production-scale continual learning. In a video from Hugging Face, viewers get an in-depth look at DeepSeek Harness (DSH), DeepSeek's open-source agent harness built on the Cordis plugin system. The core philosophy is striking in its simplicity: everything is a plugin—including the agent loop, the tools, and even the system prompt. The explainer begins with a quick tour of the new framework and unpacks how this modular design shapes the way agents are assembled and extended.

Meanwhile, the PyTorch Foundation shares a keynote from PyTorch Conference North America featuring Cody Mazza-Anthony of Shopify, who presents the company's continual learning loop built with PyTorch and vLLM. The talk describes a flywheel that transforms Shopify's product expertise into robust evaluations, then mines, critiques, and repairs low-scoring conversations to train better models. Using Buyer Profile as a case study, the keynote demonstrates how small models can deliver meaningful results within this iterative pipeline.

Together, these posts illustrate two complementary directions in modern AI engineering: highly composable agent frameworks on one side, and disciplined, evaluation-driven training loops on the other. Whether the