The Format Split That Makes LLM Agent Work Reviewable

LLM agents routinely produce plausible but broken code and documentation—work that requires careful verification before use. A hybrid approach splitting output into Markdown (for narrative, intent, and trade-offs) and TOML (for machine-checkable invariants and dependencies) enables efficient review while preserving the context needed for future maintainers. This format boundary solves a fundamental tension in agent oversight: machines need schemas to enforce correctness, but humans need prose to understand why decisions were made.