AgentMark
Markdown for AI architecture, agent topology, and volatile AI decisions. Motto: Draw the system. Record the assumptions. Expire the claims.
AgentMark is a text-first notation for describing AI systems. It begins as a terse, Mermaid-like way to sketch agent topology, and matures into an AI Architecture Knowledge Language — a single source-of-truth document that captures not only what connects to what, but why it was chosen, under what assumptions, what constraints apply, what evidence supports it, when it expires, and how we know it still works.
A diagram language answers one question: What connects to what?
AgentMark answers seven: What connects to what — why was it chosen — under what assumptions — what constraints apply — what evidence supports this — when does it expire — and how do we know it still works?
That is the leap. Traditional notations (UML, C4, BPMN, Mermaid, ADRs) document the system. AgentMark documents the system plus the assumptions, decisions, constraints, confidence levels, and the shifting AI landscape that caused the system to exist — the part AI teams lose every two weeks when the ecosystem moves underneath them.
The biggest design idea
The diagram is not the source of truth. The .agentmark document is. The diagram is just one view generated from it. One file projects into many views: topology, runtime, decision, risk, cost, eval, landscape.
The central rule
This single rule is what makes AgentMark different from Mermaid, UML, C4, BPMN, and ADRs.
Any time-sensitive statement MUST include as_of and review_by.
Any comparative statement MUST include scope and metric.
Any decision SHOULD link to claims and constraints.
Any high-risk action MUST link to policy, approval, or authority.
Any claim used by a decision MUST have confidence and evidence.
The sketch core (what stays viral)
[User] -> [Agent] -> [Model]
[Agent] -> [MCP] -> [Tool]
[Agent] -> [RAG] -> [Data]
[Agent] -> [Eval] -> [Agent]
People should never need to learn a language to start. Everything advanced is optional and layers on top.
Three modes
| Mode | Purpose | Surface |
|---|---|---|
| Sketch | Virality, whiteboard, posts | Tiny arrow grammar |
| Spec | Engineering | IDs, roles, properties, middleware |
| Decision | Time-sensitive AI architecture | Claims, constraints, decisions, invalidation |
How this spec is organized
Read the pages in order. Each is a self-contained section of the AgentMark v0.2 specification.
- Overview & Philosophy — the leap, the precedents, the three modes
- The 8-Layer Meta-Model — L1 Actors → L8 Infrastructure
- Concept Ontology & Node Taxonomy — the durable vocabulary
- Syntax & Notation — arrows, edge labels, stable IDs, preamble, SemVer
- Knowledge Layer: Claims, Constraints, Decisions — the three-way split
- Market Intelligence — landscape radar, capability matrices, tradeoffs
- Temporal Validity, Benchmarks & Invalidation — making the document alive
- Governance & Observability — trust boundaries, authority, risk, telemetry
- Tooling: Views, Linter & CLI — projections and enforcement
- Full Worked Example: Hermes Coding Harness — the complete file
- Design Principles & The Central Rule — what to avoid, what to enforce
- Grammar Specification (EBNF & Conformance) — the implementable syntax
Spec status: draft v0.2 · Owner: Joel Tong.