What Is a Harness?

Module 0.1 · The Landscape · Master Course — Harness Engineering

Senior engineer and above · 15 minutes · Prerequisite: familiarity with LLM APIs

Slides 03 · Teaching Document 01 · Diagrams 02 · Lab 07

When you say "the AI did X"…

…you are almost always describing the harness, not the model.

Everything else — the loop, the tools, the retries, the permission gates, the safety — is the harness.

The model is 1.6% of the system

MBZUAI study, May 2026. Four independently-built production agents converged on the same harness patterns. The interesting differences live in the 98.4%.

Definition

More precisely — the code, config, and runtime scaffolding that:

• LOOP runs the model repeatedly until a stop condition
• TOOLS provides named, schema-validated functions
• SAFETY enforces permission and blast-radius boundaries

Miss any one of the three → it's a prompt, not a harness.

The three jobs

while (true) {
  const r = await model.complete(messages);
  if (r.stopReason === "end_turn") return r.content;
  if (r.toolUse) {
    const out = await executeTool(r.toolUse);
    messages.push({ role: "tool", content: out });
  }
  messages.push({ role: "assistant", content: r.content });
}

The loop, as a runnable n8n workflow

Per the course's visual stack: n8n first, then Mermaid, then code. This is the same loop as the TypeScript on the previous slide, in node-graph form — importable as JSON.

       Manual Trigger
            │
            ▼
   Assemble Context ── system prompt + history + tool defs
            │
   ┌────────▼─────────┐
   │      Loop        │  ◀───────┐  stop #1: iteration cap (8)
   │  (maxIter: 8)    │          │
   └────────┬─────────┘          │
            │                    │
            ▼                    │
      Call Model ── tools: read_file, write_file, bash
            │                    │
            ▼                    │
     Stop?  ── yes ──▶ Output (done)
            │ no (tool_use)      │
            ▼                    │
     Route ─▶ read│write│bash    │
            │                    │
            ▼                    │
   Append Tool Result to History ┘  ← ONE TURN per traversal

The model is one node of seven. The turn boundary is the back-edge into Loop — the single most important concept in the course, made visible.

What a harness is not

Convergent evolution

2025–2026. Four teams, no coordination, one architecture:

The thickness spectrum

Not a quality axis. A design decision. Both ends are correct for their use case.

Every layer of thickness adds: capability, safety, observability — and also: code to maintain, failure modes, attack surface. The senior skill is defending where on the spectrum a design belongs.

The 1.6% / 67.6% trap

A typical production session. Where the context actually goes:

Tool outputs

Reading a harness from the outside in

Six steps. Same method for every harness in this course.

Lab (artifact 07): run this against Pi's source in 30 minutes — but first see the loop run in n8n.

The map ahead

Module 0.1 introduced the three-layer architecture. The next 12 modules each go deep on one layer or cross-cutting concern:

Takeaways

• The harness is the product. The model is a component.
• Three jobs: loop tools safety. Miss any one → it's a prompt.
• Thickness is a design decision, not a quality score.
• The structure converged. The variation is in thickness.
• Read any harness in six steps, outside-in.

Next: Module 0.2 — The Ecosystem. Then Module 1 — the loop, in depth.