If you have built an agent product in the recent years, you have probably written a lot of harness code: tool wrappers, context management, orchestration logic, and workflow abstractions. Even no-code tools like n8n are a form of harness engineering. The problem is that many of these abstractions exist only because of current model limitations. As models improve, parts of today's harness will become unnecessary technical debt. This seminar explores how FAVE is moving from a tightly coupled LangGraph-based architecture toward a core agent kernel with a plugin-based system, similar to the direction taken by tools such as Codex/Claude Code.