A developer intern at DiDi built agenthatch after observing that SKILL.md files, which contain agent instructions, are interpreted unreliably by large language models like Claude Code. The intern noted that strict requirements in bold text were skimmed, interpretation drifted each run, and a bug almost hit production. The core problem identified is that SKILL.md is prose pasted into a system prompt, lacking a compile step, type checking, or validation, causing skills to leak into each other within a shared context window. The intern argues skills should be source code and agents binaries, similar to how Java compiles to bytecode. Agenthatch compiles SKILL.md into real Python code that runs independently, with three commands: `pip install agenthatch`, `agenthatch hatch my-skill`, and `agenthatch run my-skill`. Compiled agents carry about 150 bytes of runtime config, schema validation happens at compile time, and the output is a real Python package. The process has three phases, with Phase 1 involving no AI, only reading frontmatter and files via pure filesystem operations.