Korva is a scroll of the internet’s best ideas.
Nothing got replaced, only layered
Context engineering
Hold the prompt constant and a model can still behave very differently, depending on the context it is working in. Context engineering covers everything the model sees: its own outputs, what the human says, and the responses from the environment. It is not about stuffing every fact into the context window. It is about picking the most relevant information, feeding it in at the right time and in the right sequence, and sometimes going back to delete things the model was told earlier so it stays on the rails.
Harness engineering
Harness engineering formalizes context into tools, files, the agent's memory over time, and parts of an enterprise's context. Claude Code and OpenAI's Codex are two well-known harnesses, and Pi is an up-and-coming open-source one. On agent benchmarks, holding the model constant and changing only the harness shifts the results by a shocking amount, even though all the intelligence sits inside the model.
Loop engineering
Loop engineering is the layer on top. A lot of productive work gets done without it, and it mostly describes what people were already trying to do: get an agent to take on more and more work over extended periods without human intervention.
Why does the harness shift benchmark scores?
The harness decides which tools a model can call, how files are shown to it, and how errors come back. A 2024 Princeton paper on SWE-agent found that redesigning that interface alone raised a model's score on the SWE-bench coding benchmark. Small details, such as how many lines of a file appear at a time, changed results. The researchers called this layer the agent-computer interface.
How did prompt engineering become its own field?
OpenAI's 2020 GPT-3 paper showed that a model could pick up a task from a few examples placed in its prompt. That made the wording and contents of the prompt a lever worth tuning. Prompt engineering grew out of that finding. In 2025, practitioners began using the term context engineering for managing everything else a model sees.
Why delete things the model was told earlier?
A model conditions on every token in its window, so an earlier wrong statement keeps influencing later output. Failed attempts and outdated instructions sit beside the current ones with equal standing. Removing them leaves the model a cleaner record to work from. Anthropic's engineers have described clearing old tool results as one technique for this.
More from Lightspeed Venture Partners
