Autonomy in Three Files: Transforming AI-Powered IDEs into Self-Directing Agent Platforms

Modern AI-powered IDEs can read files, write files, run commands, and browse the web—all the capabilities an autonomous agent needs. This paper describes a minimalist protocol that uses three Markdown files (goal.md, how-to.md, path_way.md) to transform any file-capable IDE into a self-directing agent platform. The protocol maps to a cognitive architecture: goal.md serves as long-term goal memory, how-to.md as procedural knowledge, and path_way.md as working memory. The "Next Step" field in path_way.md becomes the instruction for the next iteration—implementing recursive meta-prompting without code. Tested with Cursor and Antigravity IDEs using Gemini and Claude series models.