{"type":"video","version":"1.0","html":"<iframe src=\"https://www.loom.com/embed/7591ea06e0db482f9f87f88fef5a7e84\" frameborder=\"0\" width=\"1362\" height=\"1021\" webkitallowfullscreen mozallowfullscreen allowfullscreen></iframe>","height":1021,"width":1362,"provider_name":"Loom","provider_url":"https://www.loom.com","thumbnail_height":1021,"thumbnail_width":1362,"thumbnail_url":"https://cdn.loom.com/sessions/thumbnails/7591ea06e0db482f9f87f88fef5a7e84-c37b3b6438a2ac00.gif","duration":756.272,"title":"Crude Code Episode 1","description":"This Loom explains how to define an agent and its reusable building blocks for solving domain-specific problems with LLMs. It emphasizes that LLMs are stateless and require providing all needed context for each question, with context acting as the limiting factor. The framework includes context, connections (to external systems via tools like MCP), capabilities for domain-specific knowledge, and code to avoid bespoke logic, making agents plug-and-play and reusable. It notes model token limits such as Sonnet at 200,000 tokens and Opus 4.8 at 1,000,000 tokens, which shape design choices."}