Phase 4: Harness Engineering · 40 min · Claude Code · Python

Agent = Model + Harness

Two teams, the identical model, wildly different reliability. The difference was never the model.

Hiring signal: Describing what you actually built as 'the harness around the model' rather than 'a good prompt' is the specific language shift that signals production AI engineering experience to an interviewer.

What you will learn

Explain 'agent = model + harness' and why the harness, not the prompt, is the durable engineering deliverable
Name the six components of a production harness: rules, skills, tools, tests, feedback loops, approval gates
Predict how reliability changes when you swap the model but keep the harness, versus keep the model but strip the harness
Audit a real project for which harness components are actually present

Introduction

Agent = Model + Harness

Two teams at different companies both build customer-support triage tools using the identical underlying model. Team A's tool works well in demos and unpredictably in production — sometimes it correctly escalates an urgent ticket, sometimes it confidently misclassifies one in a way nobody notices until a customer complains. Team B's tool, same model, same general task, is reliable enough that they've stopped manually spot-checking its output. The gap isn't model quality — it's identical. Team A built a chat window with a clever system prompt. Team B built rules that constrain scope, a library of skills for edge cases, a narrowly scoped set of tools the agent can call, a test suite that gates every deployed change, a feedback loop that flags drift, and approval gates on anything customer-facing. Team B built a harness. Team A built a prompt and called it done.

The equation that matters

Agent = Model + Harness is the compact way to state something this course has been building toward since Orbit 1: the model supplies raw capability — the ability to reason, generate, and follow instructions — but capability alone doesn't produce reliable, repeatable, auditable behavior. The harness is everything around the model that turns "usually pretty good" into "reliable enough to trust." This isn't a minor implementation detail; it's the actual engineering artifact. When you swap in a stronger model six months from now, your harness — the rules, skills, tests, and gates — mostly carries over unchanged, because it encodes your product's requirements, not the specific quirks of one model version. The prompt you typed into a chat window last Tuesday carries over to nothing.

The prompt is not the deliverable
A brilliant one-off prompt that produced a great result once is not an engineering artifact — it's not durable, not testable independent of a specific session, and not something anyone else on your team can inherit or improve. The harness is durable: versioned, testable, and legible to someone who wasn't there when you wrote it.

Unlock the full lesson

You've read the first 2 sections. The rest of this lesson covers Six components, working together, What happens when you change one half of the equation, Orbit 1's tools were already harness components, Build It — plus a hands-on lab, quiz, and project artifact.

Create a free account to unlock Phase 0 and Phase 1 of every course — no credit card.

Key terms in this lesson

Agent