Glossary
What is context engineering
In short
Context engineering is the selection of everything that reaches a model before it answers: instructions, documents, conversation history, tool descriptions and memory, so that an AI agent works accurately and at low cost.
Prompts · context engineering
MMOOZZGG.com editorsChecked: 1 min read
For a child
Before a test, you have a small backpack, and not all your textbooks fit in it. You pack only what you need: the notebook on the topic, a ruler, a pen. If you stuff it with everything, you will not find the right notebook. AI also works with a small backpack of memory, and adults decide what to put in it for each task.
In plain words
A model sees only what lies in front of it at the moment it answers. This includes not just the words of the task but also found documents, past messages, search results and the list of available programs. Space is limited, and the more extra material there is, the worse the model notices the main thing. Context engineering decides what to add, what to remove and in what order. For example, a support agent gets the customer card and two recent requests, not a whole year of correspondence.
For business
For a company, this is the reason the same AI agent works in one firm and gets confused in another. Quality depends not only on the model but also on which data from the CRM, the knowledge base and the history it gets at each step. Extra text costs money, because every token is paid for, and it also lowers accuracy. What to check before launch: where the agent takes its data from, how the base is updated, what it forgets between sessions, how a long history is compressed. These decisions are made when designing an AI department, not after customer complaints.
For a specialist
Context engineering is the management of the whole set of tokens a model receives at each call: the system instruction, tools, external data, message history and memory. Anthropic, in an article of 29 September 2025, describes it as the next step after prompt engineering for agents that work over many steps, and states the goal: find the smallest set of informative tokens that raises the chance of the desired result. The reason: as length grows, retrieval accuracy falls (context rot), and a transformer has a limited attention budget. Liu et al. (2023) showed that models use information from the middle of a long context worse. Techniques: just-in-time retrieval instead of loading everything at once, history compaction, structured notes in external memory, subagents with a clean window that return short summaries, compact non-overlapping tools, and RAG for documents.
Example
An agent checks supplier invoices. Instead of the whole archive, it gets the contract with this supplier, the previous invoice and the payment rules, and it writes the results of the check into a note instead of keeping the whole conversation in memory.
Related terms
Questions and answers
Sources
Where companies use it
Not sure whether your company needs this?
AI solutions map