Agents Without Code: Skills, YAML, and Filesystems Replaced Python — Philipp Schmid, Google DeepMind

Building LLM agents is shifting from manual, code-heavy orchestration to streamlined, file-based architectures. Historically, developers managed complex Python loops, JSON schemas, and state tracking to enable agent functionality. The introduction of the Interactions API and remote agents replaces this boilerplate with isolated cloud sandboxes, enabling models to utilize general-purpose tools like Bash and CLI directly. This transition allows developers to define agent behavior through markdown files and system instructions, significantly reducing maintenance overhead. As model capabilities evolve, the focus of agent engineering is moving away from micromanaging execution paths toward defining clear domain rules and robust evaluations. This "build to delete" approach minimizes overengineering, allowing developers to leverage native model reasoning rather than maintaining fragile, custom-built harnesses.