What is context engineering?
An agent with tool access, chat history and retrieved documents can blow its context budget before it reads your actual question.
Context engineering
The practice of deciding what enters a language model's context window, in what order, and what gets dropped when space runs out. System instructions, retrieved documents, tool schemas and chat history all compete for that same fixed budget.
A simple chatbot only manages a system prompt and a conversation. An agent has more tenants: tool schemas, earlier tool results, retrieved passages, saved memory, and its own running scratchpad. Every one of those claims the same token budget.
Context engineering treats that budget as an allocation problem, not an afterthought. It decides what goes in first, what gets summarised instead of kept whole, and what gets cut once the window fills.
Why prompt engineering stopped being the hard part
Prompt engineering asks how to phrase an instruction. That question matters less every year, because newer reasoning models follow plain instructions well. What still breaks production systems sits upstream of phrasing. The right document never made it into the window. Or it made it in on turn nine, then got pushed out by turn twelve.
A bigger context window does not remove the problem. Models measurably lose track of facts buried in the middle of a long window, a pattern researchers call lost-in-the-middle. Stuffing everything in and hoping the model finds it is not a strategy. It is the failure mode this discipline exists to prevent.
- InstructionsSystem rules, fixed and small
- RetrievePull only what answers this turn
- RankOrder by relevance, not recency
- CompressSummarise history, keep decisions
- EvictDrop the lowest-value tokens first
Every turn re-runs this. An agent that never evicts anything just gets slower and less accurate as the conversation grows.
01Is context engineering the same as prompt engineering?
No. Prompt engineering writes the instruction. Context engineering decides what surrounds that instruction: which documents, which tool results, which prior turns. A well-written prompt still fails if the model never sees the fact it needs to answer with.
02Does a larger context window solve this?
Not on its own. Larger windows raise the ceiling, but models still lose track of facts placed in the middle of a long context. Every extra token also costs latency and money. A bigger window makes bad allocation cheaper to get away with, not correct.
03Where does context engineering matter most?
Multi-turn agents with tool access and retrieval. A single-shot question-answer call has little to manage. An agent running a long conversation, calling several tools and pulling from a knowledge base has a real allocation problem on every turn.
