Prompt, Context, Graph, Harness: The Way We Talk to LLMs Keeps Changing

OpenAI CEO Sam Altman said in an interview with Stratechery that he no longer thinks of the harness and the model as entirely separable things, signaling a shift in AI development toward 'harness engineering'—building a pre-wired exoskeleton around language models to enable action, memory, and self-correction. The term, popularized by engineer Vi, reflects an evolution from prompt engineering to context curation to workflow graphs, as described in the article.

Every year or so, the people building artificial intelligence decide they have been doing it wrong and give the right way a new name. The latest name is “harness engineering,” and it has an unusual endorsement behind it. In a recent interview with Stratechery’s Ben Thompson https://stratechery.com/2026/an-interview-with-openai-ceo-sam-altman-and-aws-ceo-matt-garman-about-bedrock-managed-agents/ , OpenAI CEO Sam Altman put this shift into words: “ I no longer think of the harness and the model as these entirely separable things.” That word, harness , captures a broader change in how AI systems are being built. The model is no longer treated as the entire system. It operates inside an environment of tools, context, memory, state, permissions, feedback loops, and other software that determines what it can do and how reliably it can do it. This is the idea behind harness engineering . But harness engineering did not appear out of nowhere. To understand why harness has become the word of the moment, it helps to trace that evolution: prompt, context, workflow, graph, and now the harness. In the early days of ChatGPT, the focus was entirely on how you spoke to the LLM. Engineers became “prompt whisperers,” crafting elaborate text instructions such as “You are an expert programmer,” or “Let’s think step by step” — to get the best possible output from the model. The AI system was essentially a standalone brain in a jar, and the only way to steer it was through carefully chosen words. It quickly became clear that even the smartest, fastest model couldn’t answer questions about private data or recent events it hadn’t memorized. Then in the summer of 2025, Shopify CEO Tobi Lütke and former OpenAI researcher Andrej Karpathy https://karpathy.ai/ noted that the real skill wasn’t writing a clever prompt, but curating everything the model saw. Which meant building retrieval systems to fetch relevant documents and inject them into the model’s context window. Karpathy called it “the delicate art and science of filling the context window.” https://x.com/karpathy/status/1937902205765607626 Modern AI systems don’t just see one message; they see whatever an engineer loads into memory — instructions, prior history, retrieved data. The focus shifted from how you ask to what information you provide, ensuring the model had the right facts before it spoke. But getting that loadout wrong broke the system just as surely as a bad prompt did. As AI systems grew more complex, a single prompt and a well-stocked context window were no longer enough. Tasks began to involve multiple agents, tools, decision points, and execution paths, so developers started designing the workflow itself. Using frameworks such as LangChain and LangGraph, developers began representing these systems as graphs: nodes for tasks, edges for possible transitions, and explicit rules for state, routing, retries, and human intervention. One agent might generate an answer, another critique it, and a third revise it. Unlike prompt or context engineering, the focus was no longer just on what the model sees, but on how the entire task moves from one step to the next. Note: The terminology is still evolving. “Graph engineering” is an emerging label rather than a settled discipline, but the underlying idea is clear: as agents become multi-step systems, developers increasingly have to engineer the workflow around the model rather than leave the model to determine the entire process itself. Today, the industry realizes that a raw language model is still just a stateless text predictor — it has no hands, no durable memory, and no definite way to recover from its own mistakes. Harness engineering is the practice of building a pre-wired, autonomous “exoskeleton” that wraps around the model, letting it act, fail, and self-correct in the real world. Harness engineering folds these previous ideas into a single, blunter framing. Popularized by engineer Viv Trivedy, the concept is defined by a simple equation: Agent = Model + Harness . The model handles the reasoning. Everything else — the tools it can call, the state it maintains, the guardrails it operates inside, and the evaluation loops that catch its errors — is the harness. None of these disciplines replaced the one before it so much as absorbed it. Harness engineers still write prompts, manage context, and design workflow graphs. What changed each time was the unit of engineering — from a sentence, to a window, to a network, to the whole apparatus a model lives inside. It is a fair bet that “harness engineering” will not be the last name as we continue building around LLMs either. It rarely is, in an industry that renames its own foundations about once a year. As Altman’s admission reveals, the industry spent years racing to build a bigger brain, only to discover that raw intelligence cannot operate in the real world without a reliable exoskeleton. When everyone has access to the exact same reasoning engine, the harness becomes the real product. Whatever label comes next, this one is the closest thing the field has to an admission that the model was never doing this alone. Prompt, Context, Graph, Harness: The Way We Talk to LLMs Keeps Changing https://pub.towardsai.net/prompt-context-graph-harness-the-way-we-talk-to-llms-keeps-changing-8cfe1105eac5 was originally published in Towards AI https://pub.towardsai.net on Medium, where people are continuing the conversation by highlighting and responding to this story.