Introducing melchizedek
A capable model is not a capable system. Give a frontier model a vague goal and broad tools and you get brilliance on Monday, nonsense on Tuesday, and no way to know which is coming. The unlock isn’t a smarter model. It’s coordination. Melchizedek is my orchestration framework built on that conviction.
agents as defined work
The first discipline is narrowing. An agent with one job, a bounded tool
set, and explicit output rules produces predictable outcomes — not because
it’s less intelligent, but because it has fewer ways to be wrong. A
research agent that must cite tool results instead of memory. An analyst
that must return exact figures or say DATA_NOT_FOUND — never guess. A
critic that must answer in structured JSON, every time.
In melchizedek, that definition is the artifact. An agent is a YAML block: name, model, instruction, tools. Nothing hides in code. You can read a subagent’s entire contract in thirty seconds, diff two versions of it like config, and swap its underlying model — Gemini to Claude — by changing one line. Constraint, made legible.
orchestration as the factory
The second discipline is the line between the stations. A single do-everything agent is a craftsman: gifted, variable, unaccountable. A syndicate — melchizedek’s word for a hierarchical agent graph — is a factory: work moves through defined hand-offs, and the structure itself enforces quality.
The patterns are old manufacturing, reborn. A delegation router reads each request and routes it to the specialist — code to the coder, math to the mathematician. A critic loop is the QA station: one agent drafts, another inspects against a rubric, the draft doesn’t ship until it passes. A hierarchical decomposer is the foreman, splitting a goal into subtasks and reassembling the results. A council fans one question out to parallel analysts and synthesizes their reports into a single view.
Watch the routing pattern run — one request, two hand-offs, zero improvisation:
None of these patterns require clever prompting. They require placement — which agent stands where, who hands off to whom, who is allowed to say no. The whole factory floor fits in one YAML file, and that is the point: the architecture of delegation should be something you can read, version, and reason about, not something buried in a call stack.
Every post in this site’s foundations thread will keep returning to this: coordination is the discipline that makes intelligence composable. Coordination is how a system becomes more reliable than any of its parts. Coordination is the difference between a talented mind and a working institution — and we are, all of us, about to become managers of institutions we cannot see.