What is context engineering?
The discipline of deciding what information an AI model gets and how: the successor to prompt engineering for serious systems.
Context engineering is the discipline of designing everything an AI model sees before it answers: instructions, retrieved documents, history, data from business systems, definitions of the tools it can use. The term took hold between 2025 and 2026 (pushed among others by Andrej Karpathy) as the evolution of prompt engineering: no longer writing a good question, but building the entire information environment the model operates in. The final answer's quality therefore depends as much on what gets placed in that context as on the model chosen: which documents to retrieve, in what order to present them, what to summarize and what to discard become design decisions on par with writing the prompt itself. For this reason context engineering today involves not just whoever writes prompts but also whoever designs the data pipelines feeding the model.
Why it replaced prompt engineering
As long as AI was a chat, phrasing the request well was enough. In real systems (assistants over company data, agents executing processes) output quality depends far more on what is in the context than on the phrasing: which documents were retrieved, in what order, what was summarized, what was left out. The context window is a scarce resource: filling it badly means worse answers and higher costs. The prompt remains, but it has become the small part of an engineering job.
What it means in practice
In an enterprise project, context engineering means deciding how to chunk and index documents for RAG, which structured data to inject and from which systems, how to compress conversation history, how to describe tools to an agent. It is work on data before work on models: the reason successful AI projects almost always start by putting the sources in order, not by choosing the LLM. The direction agentic systems are moving in is just-in-time retrieval: instead of preloading everything that might be needed, you give the system lightweight references and the tools to fetch the right piece at the moment it is needed, often combined with a share of context loaded up front.
Related terms
- RAG · A technique that grounds an LLM in your company data: it retrieves relevant documents and feeds them to the model before it answers.
- AI Agents (Agentic AI) · AI systems that go beyond answering: they plan, use tools and take actions autonomously inside your processes.
- Knowledge graph · A network representing company data as entities and relationships: customers, products, contracts and the links connecting them.
- Context Rot · The decline in an LLM's relevance and accuracy as its context fills up, well before it hits the stated limit.
- Intent engineering · The discipline of defining an AI agent's goals, constraints and stop rules in production, beyond any single prompt or turn.
A term that hits close to home? Let's talk.
CONTACT ME