Your Data. Your Device. Your Software.

Humans read linearly. Agents query.

A retrieval model for how software agents consume documentation, with a Markdown binding. A human reading top-to-bottom resolves “as discussed above” from continuous context; an agent handed one section as a retrieved chunk cannot. AgenticMD makes every section safe to retrieve in isolation — a stricter profile of CommonMark in the lineage of Pandoc Markdown and MyST, with no new file format to learn. v0.6 made the per-type retrieval rules normative on runtimes; v0.8 + v0.9 added the measurement substrate so adopters can publish comparable benchmark data, not just architecture claims.

• Typed nodes Every section carries a marker drawn from a closed ten-value vocabulary. Agents route on the type — guardrails read differently from architecture notes, corrections invalidate prior claims, untrusted content gets gated before model context.
• Addressable references Cross-references resolve by ID against the corpus, not by file path. Reorganize, merge, or rename — the links survive. Edge typing (operational vs semantic) lets agents walk the right subset for the current intent.
• Brief-first probes Every topic opens with an 80-word brief. An agent scans twenty briefs for less context than reading one full topic, then pulls deeper under a declared disclosure budget that the runtime is normatively required to honor.
• Measurable v0.8 ships a metrics taxonomy and JSON-Lines telemetry schema; v0.9 ships the reference library, canonical task suite, and benchmark corpus. Adopters can publish comparable retrieval data — the substrate for empirical claims, not just architectural ones.