The curriculum opens
This post records the course at its 2026-07-11 launch. The curriculum has since grown into three parts: LLM fundamentals, taught from open-weight local models down to the mechanism underneath; agent design, from single-agent anatomy to the full melchizedek protocol with memory and MCP; and agent observability, a third part in preparation. The syllabus is current; the account below is historical.
You learn agent architecture by building it. The curriculum puts working syndicates between the essays and your own machine, in six modules. Five are built on real agent teams from the Melchizedek repository: you read their YAML, follow their traces, and work their logic by hand. The sixth takes apart the agentic coding loop that built the course itself. Progress is kept in your own browser, and the credential is a system you can rebuild without us.
The exercises are built the way the course teaches. Each one is a schema-validated content file rendered by a shared family of components: terminal checkpoints, a prompt forge, a topology router, a memory triage bench. The debugging panel runs on outputs from runs that actually failed. A wrong answer returns an amber marker and an authored explanation of why it was wrong, because a checkpoint that only scores you has taught you nothing.
The specimens came from live runs. The second module put one question through the Style Council’s three engineered voices, a system built for the course. The third module’s failing prompts are historical runs, all five revisions, two of which genuinely fail. The fifth module ran a whole visual pipeline in the open: generate an image, inventory it blind, audit the inventory against the approved specification. Every syndicate ships as a download with rebuild notes. Melchizedek is the specimen; the patterns move to any stack you work in.
Here is one turn from the live council run:
Begin at the syllabus or type curriculum in the shell below.