Insights · point of view
The work moved from wording a message, to choosing what the model sees, to building the machine around it. A short field guide to three eras — and why the harness is where outcomes are won now.
IANIA.AI · 8 min read
Pick two frontier models today and run them against most enterprise tasks. The gap between them is narrow and shrinking. That is the uncomfortable truth behind a lot of AI strategy: the model is no longer the differentiator. What separates a useful system from an impressive demo is everything you build around the model — and that has changed shape three times in three years.
The first craft was wording. You packed the role, the instructions, the examples and the output format into a single message and hoped it landed. Get the words right, get a good answer. There was no memory of past turns, no tools, no system around the call. When it failed, you started over and rephrased.
This still matters — clear instructions always will — but as a discipline it tops out quickly. A perfectly worded prompt over the wrong information, with no ability to act, is a confident answer to a question nobody can verify.
Named by Andrej Karpathy, context engineering reframed the job. The model only ever sees a finite window. The real work is deciding what goes into that window and what gets cut: pulling from documents, tools, memory and previous turns, then curating, compressing and dropping to fit. The unit of work stops being the sentence and becomes the state — what the model can see at each step.
The lesson is blunt: better context beats a cleverer prompt. Most “the AI got it wrong” failures in production are not reasoning failures — they are context failures. The model was never shown the thing it needed.
Named by Mitchell Hashimoto, harness engineering is the current frontier. It keeps everything context engineering does and adds the scaffolding bolted onto the model: personalisation, persistent memory, tools and the ability to act, and delegation across steps and sub-agents. Strip those away and you have a chatbot. Bolt them on, with evaluation and guardrails, and you have something that runs a job end to end.
This is the shift that matters for organizations. “Make the AI better” stops being a prompting task and becomes an engineering one: retrieval and knowledge systems, tool integration, memory, orchestration, evaluation, observability and governance — designed together. The harness is the product.
Three practical consequences. First, stop benchmarking vendors on the model alone — benchmark them on the harness they can build and operate. Second, treat context as architecture: your knowledge systems, retrieval and data are now first-class engineering, not an afterthought. Third, budget for the unglamorous parts — evaluation, observability and governance are what let a harness run unattended without becoming a liability.
None of this means prompts stopped mattering. It means the leverage moved. In 2024 the lever was the wording. In 2026 it is the system — the harness you design, ground in your knowledge, and govern.
That system is what we build. See how →
The terms context engineering and harness engineering were coined by Andrej Karpathy and Mitchell Hashimoto respectively. The framing and diagram here are our own.