Harness engineering: why the value of AI has left the model

The whole industry stares at the same scoreboard: which model is in the lead, by how much, on which benchmark. The reflex is understandable, but it misses the point. The best models now sit within a few decimals of one another. While we compared those decimals, the engineering that actually decides things moved to new ground: it shifted from the model to what we build around it, the harness.

I sum this up with an equation that has become my compass: Agent = Model + Harness². The model counts, of course: it is one of the two terms, and it sets the ceiling of what the agent can reach. But the harness is squared. It is the entire environment that turns raw intelligence into a genuinely useful agent: its tools, its memory, the business procedures it has to follow, its guardrails, its verification loops, its access to the information system. Two agents built on the same model can end up worlds apart depending on the quality of that second term. More often than not, it is the term that makes the difference.

Three ages of AI engineering

To understand this shift, you have to place it inside a short history in three acts.

2022-2023, prompt engineering. The art of whispering into a model's ear. You phrase the right request, with the right role and the right examples. It was new, spectacular, and already a little fleeting.

2024-2025, context engineering. A good question is no longer enough, you also need the right context. Not "write an email", but "here is the client, here is our tone, here is the history, write an email". You no longer tune the phrasing, you select the information that enters the model's window.

2025-2026, harness engineering. You no longer just tune a prompt or a context, you build the whole environment around the AI. This is what takes you from a raw model to a production agent. My rule of thumb: a good prompt counts for 10% of the result, the right context adds 30%, and the harness wrapping the whole thing counts for the remaining 60%.

The proof in the numbers

In March 2026, the LangChain team moved its coding agent from 30th to 5th place on the Terminal Bench 2.0 benchmark without touching the underlying model: with the model held fixed, the score climbs from 52.8% to 66.5%, purely by reworking the harness.¹ On SWE-bench Pro, the finding is of the same order: swapping the harness shifts the score by more than 22 points, while switching to the best model of the moment shifts it by about one.²

Read those figures a second time. With the model held constant, the harness makes twenty times more difference than the model itself does with the harness held constant. That is the whole thesis: the model is becoming a commodity, and the variable that decides production performance is the harness.

What is really inside a harness

A good harness is an architecture in its own right, far more than a configuration file. And it is not just about code: the same layers structure a customer-support agent, a legal assistant or a financial agent. You always find tool orchestration (how the agent chooses, chains and executes its actions, with error recovery: querying a CRM, sending an email, booking a slot, calling a payment API); memory and context (indexing a knowledge base, a customer's history, contract archives); the business procedures it has to respect (the path of a refund, a compliance policy, the brand's tone); the guardrails (strict bounds, budget caps, human validation before a transfer or a message to a client); the verification loops (checking an answer against its source, validating a figure, reviewing before acting); and observability (telemetry, tracing, audit logs).

The right image is organizational before it is technical. Building a harness is like setting up a company: once you have recruited the right talent, you still have to build the organization, the tools and the processes so they move in the same direction and create value. The model is a superhuman worker. The harness engineer is the architect who decides which materials that worker is allowed to use, and the order in which the stones must be laid.

Value migrates

This shift has very concrete consequences for who makes money, and how. Stripe merges more than 1,300 code changes per week generated by its agents, without a single line written by hand.³ There is no miracle model in there: its engineers spent months building a harness that produces code consistent with the company's specification and workflow.

Conversely, platforms like Lovable have watched their advantage erode. Their real value was precisely the harness they built around the models. But that harness is becoming a commodity. An ecosystem like Claude Code now ships with its own built in: file read and write tools, permission controls, correction loops. On top of that, an open-source plugin like Hypervibe, which I built, adds everything needed to build and deploy a complete web application: database, authentication, payments, a custom domain, going live. Claude Code plus Hypervibe thus deliver natively what these platforms used to sell, and their advantage melts away. Value then simply migrates from the model to the way you wire it into your processes.

A durable engineering discipline

One last difference, and it is decisive. Prompt engineering had something of a trick to it: a good phrasing, quickly copied, quickly outdated. Harness engineering is of a different nature. It is a real craft, a real engineering discipline, one that compounds and refines over time. And the more capable the models become, the more this shift intensifies: their intelligence becomes abundant, and what stays rare is the way to wire it into a real context.

That is also what makes the harness so strategic. Done well, it lets you make AI the operating system of your organization. I am building one right now for my own activities, and it has transformed close to 90% of the way I work. The term itself is young, it emerged in early 2026, but the skill will last.

The conclusion fits in one sentence. In the era now opening, the reigning skill is shifting: knowing how to build the harness around the model. Stop looking for the right whisper. Become the architect.

Footnotes

1. LangChain, "Improving Deep Agents with harness engineering" (March 2026). With the model held fixed (GPT-5.2-Codex), the agent goes from 52.8% to 66.5% on Terminal Bench 2.0, from top 30 to top 5. https://www.langchain.com/blog/improving-deep-agents-with-harness-engineering ↩

2. "Agent Scaffolding Beats Model Upgrades" (Particula, 2026). On SWE-bench Pro, the scaffold accounts for a swing of more than 22 points, where switching to the best current model moves it by only about one. https://particula.tech/blog/agent-scaffolding-beats-model-upgrades-swe-bench ↩

3. "Stripe's Autonomous Coding Agents" (InfoQ, March 2026). Stripe merges more than 1,300 pull requests per week produced by its agents, with no hand-written code, each reviewed by a human. https://www.infoq.com/news/2026/03/stripe-autonomous-coding-agents/ ↩