Prompt Engineering Is Dead. Long Live Harness Engineering.

I spent 3 months perfecting prompts. Then I deleted half of them.

In late 2023 I had a directory called prompts/ with 47 carefully tuned templates. Few-shot examples, Chain-of-Thought scaffolds, a tiny ReAct loop I was very proud of. I'd A/B tested wording. I'd argued on Twitter about whether "Let's think step by step" still worked.

By mid-2025 I deleted 23 of them. They weren't wrong. They just weren't the bottleneck anymore.

The thing that broke my agents in production was never the prompt. It was the environment around the prompt — the tools they could call, the files they could see, the moment the loop should stop, the rollback when a tool returned garbage. The prompt was a polished doorknob on a house with no foundation.

That's the story of the last three years of AI engineering, compressed: we keep renaming the layer where the real problem lives.

A 40% failure rate, and it's not the model's fault

Here is the number that should embarrass us. In 2026, around 40% of AI agent projects fail in production. Y Combinator's DevTool Day surveyed CTOs and CPOs in March 2026 and found a strikingly consistent post-mortem: "the difference between success and failure isn't the model."

75% of YC enterprise companies have already deployed coding agents. Most of them hit the same wall: the demo works, the prod deploy collapses. Linear declared in March 2026 that "issue tracking is dead" — meaning if your coding agent gets the issue context directly, you don't need a human ticketing layer at all. Enterprise workflows are being redesigned around agents.

In that environment, shipping an agent without understanding the harness around it is like merging onto a highway without a seatbelt. You'll go fast. You'll go through the windshield on the first curve.

So how did we get here? Three stages.

Stage 1: Prompt Engineering (2022–2023)

Scope: one input string.

Prompt engineering optimized a single message. Few-shot examples. Chain-of-Thought. ReAct. The deliverable was the wording itself, and we treated it like poetry. Some of it was poetry. A lot of it was incantation.

What it solved: a single LLM call going from 60% useful to 85% useful, on a single task, with no tools and no loop.

What it couldn't solve: anything that needed the model to do something rather than say something. The moment you wanted the model to call a function, read a file, remember yesterday's conversation, or decide between three branches, your beautifully tuned prompt was a single-cell organism trying to run a marathon.

I remember the exact week I realized this. I'd written a prompt that scored 92% on my eval set and 11% on real customer tickets. The eval set didn't have the messy attached PDFs, the half-deleted Slack thread, the customer who said "you know, the thing." The prompt was perfect for an environment that didn't exist.

Stage 2: Context Engineering (2024–2025)

Scope: everything the model sees at inference time.

Andrej Karpathy reframed it: "it is a lot more than just the prompt itself." Philip Schmid (formerly Hugging Face, now Google DeepMind) was blunter: "the new skill in working with AI is not prompting, it is context engineering."

Context engineering treats the entire context window as the artifact. System prompt, retrieved documents (RAG), tool definitions, conversation memory, structured outputs from previous turns — all of it. Anthropic's own framing in their Effective Context Engineering post calls it "the natural progression of prompt engineering": you're still curating tokens, you're just curating a lot more of them, and most of them weren't typed by a human.

This is when MCP showed up, when vector databases stopped being a curiosity, when "what does the agent know right now?" became an actual debugging question with an actual answer.

What context engineering solved: agents that could read your codebase, recall a meeting from last week, and call a real API with the right schema.

What it didn't solve: what happens when that agent runs for six hours unsupervised, calls 200 tools, accidentally rm -rfs a sandbox, and there's no one watching. Context engineering tells you what the model sees. It doesn't tell you what happens when the model acts and something goes sideways.

I learned this the expensive way when an agent of mine, with a beautifully engineered context, spent four hours and $38 in tokens recursively summarizing its own summaries because nothing in the system told it to stop. The context was perfect. The environment was a fire.

Stage 3: Harness Engineering (2025–)

Scope: the entire environment the model operates inside.

Context + constraints + tools + lifecycle + feedback + observability.

Louis Bouchard puts it cleanly: "Context engineering is what you send to the model. Harness engineering is how the whole thing runs."

If prompts are the recipe and context is the ingredients, the harness is the kitchen — the fire suppression, the timers, the knife rack out of the toddler's reach, the voice that says "chef, table 4 just walked out."

Concretely, a harness includes:

Constraints — what tools the agent is not allowed to call, what files it can't touch, what it must escalate to a human.
Lifecycle — when does a task start, branch, retry, give up, hand off?
Feedback loops — how does the agent know its last action worked? Linter? Test run? User reaction? Another agent grading it?
Observability — when it fails at 3am, can you reconstruct why in under 10 minutes?
Permissions and isolation — can it write to prod, or only to a worktree?

The term went mainstream in early 2026 for a specific reason: every major lab made the same architectural bet at once. Anthropic shipped Managed Agents in April 2026 at $0.08 per session hour and introduced a three-agent harness separating planning, generation, and evaluation for long-running coding work. OpenAI updated its open-source Agents SDK with a model-native harness. Google and Microsoft followed. The New Stack summarized the moment with a headline I won't forget: "They agree the harness is the product. They disagree on the price."

When four labs that disagree on everything agree the harness is where the value is, that's a signal.

Replace, or stack?

There's a real disagreement in the field about whether harness engineering replaces prompt and context engineering or contains them.

The replace camp (Data Science Dojo, others) argues that the world agents now operate in wasn't anticipated by the prompt-and-context era, so we should retire the old vocabulary and start clean.

The stack camp (AnyTech, several others) argues there's no fundamental difference — the vocabulary just keeps growing because LLMs keep doing more, and we shouldn't throw away knowledge every time a new buzzword arrives.

I'm in the stack camp, with one caveat: stacking only makes sense if you understand the containment relationship.

Harness ⊇ Context ⊇ Prompt

Prompt engineering is still essential. A bad prompt inside a great harness still produces bad output. Context engineering is still essential. Garbage retrieved documents poison the smartest harness. But by 2026, prompt-only and context-only thinking has run out of altitude. The interesting bugs — the 40%-failure-rate bugs — live in the outer layer.

The reason is unromantic. Agents got long-running. Once an agent operates autonomously across hours and hundreds of tool calls, no single prompt can steer it and no static context can describe its world. The environment becomes the load-bearing thing.

Why it matters now (and not in 2024)

A Dev.to post from WonderLab put it well: "The timing isn't a coincidence. In 2025, AI agents went from 'fun demo' to 'actual productivity tool.'"

Demos forgive a lot. A demo agent runs for 30 seconds, in front of a sympathetic audience, on a path the demoer has walked five times. A production agent runs for hours, on inputs nobody anticipated, while you're asleep. The forgiving environment is exactly what made prompt engineering feel sufficient. The unforgiving environment is what made harness engineering necessary.

This is also why "harness engineering" suddenly has multiple competing definitions from multiple companies — I wrote a separate piece about five companies and five definitions of the term. Everyone agrees the layer matters. Nobody agrees yet on its boundaries. That's how new disciplines look in year one.

Where I see this going

Honestly? I expect the word "harness" to get embarrassing within 18 months. We'll either have absorbed it into "agent engineering" (the umbrella term gaining ground), or split it into four more specialized terms (orchestration engineering, eval engineering, permission engineering, runtime engineering — pick your poison).

The vocabulary will keep moving. The underlying problem won't. The problem is, and has always been: we are putting probabilistic systems in environments that punish probabilistic behavior, and the environment is the part we keep forgetting to design.

I'm fine being wrong about the term. I'd rather be right about the layer.

And if you're still maintaining a prompts/ directory of 47 templates, no judgment. I have one too. It's just a lot smaller now, and most of the file is comments explaining what the harness around it does.