What is context engineering?

Context engineering decides, for every call to a language model, which tokens it gets to read. Those tokens are the system instructions and the tool definitions. They are also the records a search returned, earlier tool results, notes the model kept and the conversation so far.

Prompt engineering is one part of it: writing the instruction well. Retrieval is another part: finding the records that might matter. Context engineering is the decision between them. Which of those records go in, in what form, next to what, and which earlier material is taken out again to make room.

The term spread in June 2025. Simon Willison collected the posts that started it. Shopify’s Tobi Lutke liked that it names “the art of providing all the context for the task to be plausibly solvable by the LLM.”

The change in name follows a change in the work. A prompt for a single answer is written once. An agent runs in a loop: it calls a tool, reads the result and calls another, and each turn adds to what the model sees. The question stops being how to phrase the request. It becomes what the model should be looking at on turn 12.

Anthropic’s engineering team describes context engineering as what prompt engineering grows into once models work over many turns. Its guiding rule is to “find the smallest set of high-signal tokens that maximize the likelihood of your desired outcome.”

Every model has a context window: a limit, in tokens, on what one call can read. Everything below shares it, on every call:

• Instructions: the system prompt, and the rules that hold for this task.
• Tool definitions: each tool’s name, description and parameters, sent whether or not the tool is used.
• Retrieved records: tickets, part records, supplier notices, pages of a manual.
• Tool results: what each call returned, from the first turn on.
• Memory: notes the model wrote earlier, or facts kept from past conversations.
• History: the person’s messages and the model’s own replies.

Each of these is a decision someone made, or forgot to make. A tool result left in the history after it was used still takes room on every later call.

Context windows now reach millions of tokens, so it is tempting to send everything. Chroma tested 18 models in July 2025 on tasks kept simple on purpose, changing only the length of the input. In one test the models answered questions about a long chat history. Given only the relevant parts, about 300 tokens, every model did markedly better than with the full history of about 113,000 tokens. The report found that performance “grows increasingly unreliable as input length grows.”

Position matters too. Liu and coauthors showed in 2023 that models use information best at the start or end of a long input. They use it worst in the middle. Anthropic’s post calls the general effect context rot: every added token draws on the model’s limited attention.

There is a plain cost as well. Every token in the window is sent again on every call of the loop, so a crowded window costs something on every turn.

• Retrieve late, by reference. Keep ids, paths and queries in context, and let the model fetch a record when it needs it. Anthropic calls this just in time context.
• Return less from tools. A tool that returns ids and a few fields leaves room. A tool that returns a whole document fills the window with text the model may never use.
• Disclose step by step. Give the model names and one line descriptions up front, and the full text only when it asks.
• Compact long runs. Near the limit, clear old tool results and replace the history with a summary. Keep the latest result.
• Keep notes outside the window. The model writes progress to a file or a record and reads it back after a reset.
• Split the work. A sub-agent explores in a window of its own and hands back a short summary. Anthropic reports those summaries are often 1,000 to 2,000 tokens.

• Everything goes in. The answer was in the window, in the middle of 40 near matches, and the model picked a near match.
• Records the person may not open. Whatever is in the window can end up in the answer. The access check belongs before a record is added, which is what permission-aware retrieval means.
• Text read as orders. A ticket or a file in the window can carry instructions, and the model reads them like any other text. That is prompt injection.
• Stale memory. A note kept from last month states a fact that has since changed, with nothing to say when it was written.
• Compaction that loses the point. A summary drops the one detail that mattered three turns later.
• Overlapping tools. Two tools that do nearly the same thing cost room in every call and leave the model unsure which to use.

Six questions, asked of a real run rather than a demo:

1. Can you see the exact context of any call: every instruction, record and tool result the model read?
2. How much of that context did the answer use?
3. Was every record in it one the person asking may open, checked before it went in?
4. What happens at the limit: an error, silent truncation, or a summary you can read?
5. Do tool results return ids and short fields, or whole documents?
6. Does every remembered fact say where it came from and when?

The answers depend on the data underneath. A model can only be handed the right three records if something already knows which records relate to the question. That is the job of a context graph.

Curiosity Studio starts from records connected in one typed graph, from more than 70 systems, with the access rights of their source. The built in search and read tools run as the person who asked, so a record they may not open never reaches the window.

Skills follow the step by step pattern. The documentation describes how an assistant is told each skill’s name and one line description up front. It reads the body only when it needs it. The agent loop guide suggests tools that return ids and short fields rather than whole documents, and fewer than about 10 tools per agent.

The model is a setting, from any of the providers Studio supports. Each provider declares its context window. The chat uses it to decide how much history and context to send, as the custom provider page describes. The AI agents page has the developer view.