First, an Experiment

In early 2026, developer Can Bölük ran an experiment: same model, same set of tasks, nothing changed except the engineering layer around the model—specifically, the format used by the harness to handle code patches. The task success rate jumped from 6.7% to 68.3%.

Tenfold. Not a single line of the model changed.

That number has been making the rounds in AI circles lately because it turns something many people had vaguely sensed into a conclusion that is impossible to ignore: at this point, what creates the gap is no longer the model itself, but the layer wrapped around it.

That layer is called the harness—the set of equipment that lets a horse pull in the right direction. Looking back over the past few years, the AI world’s collective focus has actually shifted three times.

2022 to 2024: Everyone Learned How to “Talk”

In the first two years after ChatGPT came out, the whole industry was obsessed with Prompt Engineering.

Back then, models were dumb. And dumb models meant wording could make a real difference: add “You are a senior expert” at the beginning, and the answer would look a little better; tell it to “think step by step,” and accuracy really did improve. Prompting playbooks were everywhere, “prompt engineer” became a new job title, and for a while, courses teaching people how to write incantations were more expensive than programming courses.

Looking back now, the essence of that whole era was: treating the model like an oracle and studying how to make wishes. The thing being optimized was a single input-output exchange, and the bet was that magic could be hidden inside one sentence.

2025: The Model Wasn’t Dumb—You Were Feeding It Wrong

In 2025, the wind shifted. With coding agents like Claude Code, models started working over longer periods and across multiple steps, and a new problem became obvious: often, the model didn’t fail because it didn’t know how. It failed because it never saw the information it needed—the relevant code wasn’t included, the project conventions weren’t explained, and it didn’t remember what had been done last week.

So the focus moved from “write one good sentence” to “build a good information environment”: RAG, memory systems, the MCP protocol, project instruction files like CLAUDE.md—all of these became hot topics that year. Shopify’s CEO and Karpathy helped popularize a new term one after the other: Context Engineering—within a limited context window, deciding what to include, what to leave out, and in what order is a craft.

My own toolkit grew out of this period: memory banks, working discipline, context hygiene—all of it belongs to this school.

2026: Once Single Conversations Hit Their Limit, the Outer Layer Takes Over

But context engineering only optimizes “one conversation.” Real work is not one conversation; it is a chain: tasks need to be broken down, code needs to be written, tests need to run, failures need retries, and finished work needs review. The engineering structure that strings all of this together—tools, permissions, hooks, scheduling, review, pipelines—is the harness. This year, people have realized that optimizing a single conversation has a ceiling, but harnesses do not.

A few symbolic examples:

Stripe’s internal coding agent is called Minions. It can be triggered in Slack with an emoji reaction, and it now produces more than 1,300 PRs per week. Not a single line of code in those PRs is written by humans; engineers only do review. It runs on a heavily customized open-source harness based on Goose, with internal workflows packaged into something called blueprints. Stripe’s own breakdown is: the system is 60% model and 40% harness engineering—and that 40% is the part other companies cannot copy.

Someone dissected the source code of Claude Code and found that the part truly responsible for “calling the model” is only a small fraction. The overwhelming majority is tool definitions, permission control, context compression, and task scheduling—all harness. The AI tool you use every day is itself half a harness.

Cole Medin went a step further and turned this into a product: his open-source Archon calls itself a “harness builder,” packaging your entire agentic workflow into YAML workflows that can be orchestrated, reused, and repeated. The slogan is blunt: make AI programming less like a coin toss and more like a production line.

Even benchmarks are catching up: this year’s coding-agent evaluations have started testing the full “model + harness” stack instead of evaluating bare models in isolation.

Why This Hits Home for Me

Because everything I’ve been doing over the past two months was, in hindsight, harness work—even though I didn’t know the word at the time.

The watchdog llmdog manages whether a dozen-plus services live or die; cron pipelines run in three shifts a day; TG push notifications need token-bucket rate limiting; parallel sessions fight over git, so rules have to be established; no new project is allowed to start until the previous one is wrapped up. Not a single line of this is “model capability,” but without it, model capability is worth nothing.

Last month, I wrote a piece called “AI’s Development Curve,” where I said that everything that cannot fit into the context becomes a breeding ground for hidden bugs. Back then, I described the illness but did not prescribe the cure. Now it looks like the industry’s answer is the harness: the model’s field of view is local, so use engineering structure to contain the locality—review contains quality, testing contains regressions, pipelines contain process, and watchdogs contain runtime.

Finally

Put simply, the past four years have been a step-by-step process of admitting reality: first, we admitted that models cannot read minds, so we learned prompting; then we admitted that models cannot see our projects, so we managed context; finally, we admitted that models cannot get things done alone, so we built harnesses.

The lesson for ordinary people is right here. Making models stronger is the lab’s job; you have no say in it. But the harness is your job, and every line of it is under your control. Over these four years, the point of leverage has kept moving: from “Can you ask good questions?” to “Can you provide the right information?” and then to “Can you build a system?”

It is no accident that crash courses in prompt engineering have all gone out of fashion.