What is Context engineering?

How it works

A model only knows what is in its context window for the current request. In a simple chat that is a system prompt and a few messages. In an AI agent that runs for dozens of steps, the context fills with tool results, file contents and old turns. Anthropic's engineering guide describes context as a finite resource with diminishing returns: the more tokens you add, the harder it becomes for the model to use all of them well.

Context engineering is the set of choices that manage this:

• Select: retrieve only the relevant documents (RAG) instead of pasting everything.
• Load just in time: give the agent tools to look things up (search, read a file) rather than preloading data it may not need.
• Compact: summarise long histories and drop stale tool output once it has been used.
• Remember outside the window: write notes or task lists to files or a memory store and read them back later.
• Isolate: hand focused sub-tasks to sub-agents with their own clean context and take back only a summary.
• Keep tools lean: a small set of clear, non-overlapping tools costs fewer tokens and causes fewer wrong calls.

Why it matters

Most agent failures are not the model being "not smart enough". They come from missing information, contradictory instructions, or a context so full of noise that the key fact gets lost. Large context windows do not remove the problem; they make it cheaper to be careless. Good context engineering lowers cost and latency and makes long tasks finish reliably.

Example

A coding agent is asked to fix a bug. A naive setup pastes the whole repository into the prompt. A context-engineered setup gives the agent a short project guide, search and file-read tools, and a rule to keep a running notes file. The agent searches for the failing function, reads three relevant files, fixes the bug, and summarises earlier steps when the history grows long.