What is context engineering?

A model can only reason over what is in its context window - the text, tools, and data you place in front of it at inference time. Context engineering is the discipline of assembling that window deliberately: choosing what to include, what to leave out, and how to arrange it so the model has exactly what the task needs and little else.

The term was popularized in 2025 by practitioners across the field who argued that the hard part of building with LLMs had shifted. It is less about the wording of one prompt and more about the system that feeds the model the right context, at the right time, on every turn.

Prompt engineering is about how you phrase a single instruction - the wording, examples, and formatting of one message. It still matters, but it treats the prompt as the whole input.

Context engineering zooms out to the entire window. The prompt is just one component; the rest is retrieved knowledge, tool definitions, prior messages, and memory that a system assembles dynamically for each step. In an agent that runs dozens of turns, no static prompt can anticipate what each step needs - so the work moves from writing a prompt to engineering the pipeline that builds the context.

A useful way to think about context engineering is as a budget: the window is finite, and every token you spend should earn its place. The main things competing for that space are:

More context is not better context. A bloated window is slower, costs more, and - counterintuitively - produces worse answers, because models struggle to find the relevant detail buried in a wall of text. The failures tend to look like this:

Good context engineering is mostly about curation and freshness, not volume. A few principles hold across most systems: