Context Engineering
What context engineering means in practice
A model’s output is only as good as what it has in front of it at the moment it responds. Context engineering is the work of assembling that: which documents to retrieve, which prior conversation history to include, which system instructions apply, and how much of each fits within the model’s context window without crowding out what matters most.
In a professional services setting, this might mean an agent reviewing a contract needs the current clause, the firm’s standard playbook for that clause type, and precedent from prior similar contracts, but not the entire client file. Deciding what to include and what to leave out, and in what order, has a direct effect on the quality and reliability of the output.
How it differs from prompt engineering
Prompt engineering is about how an instruction is worded. Context engineering is about what information surrounds that instruction: the retrieved documents, the conversation history, the tool outputs, the system rules. As agentic systems have matured, context engineering has become the more consequential discipline of the two, because a well-worded prompt still fails if the model is missing the information it needs, or is given so much irrelevant information that the relevant part gets lost.
Governance note
Context engineering has a governance dimension that is easy to overlook: what a model is given access to at the point of reasoning is also what it could expose or act on incorrectly. Retrieval systems that pull more than a task requires increase both the risk of an incorrect output and the risk of an inappropriate disclosure. Designing context deliberately, rather than defaulting to “give it everything”, is as much a data governance decision as a technical one, and sits naturally alongside data backbone design.
Insight
Misconceptions
Context engineering is not simply "using a bigger context window". A larger window does not resolve the retrieval problem; irrelevant or poorly ranked information still degrades output quality even if it technically fits. It is also not a one-off setup task. As workflows and data sources evolve, what counts as the right context for a given task changes with them, which is why context design is treated as part of the Blueprint and Specification stages rather than a fixed technical decision made once and left alone.