What Is Context Engineering, and Why Does an Agent Need a Backpack?

Opening Scene

The night before a multi-day trek, an experienced guide lays every candidate item out on the floor of the lodge: rope, extra socks, a water filter, a satellite phone, three days of food, a first-aid kit twice the size it needs to be. Nothing gets zipped into the pack yet. First comes the harder work — deciding, item by item, what this specific route, this specific weather, this specific group of travelers actually requires, and what simply doesn’t earn its place on someone’s back for the next four days.

In Plain English

Context engineering is the discipline of deciding exactly what information an AI agent receives before and during a task — the instructions, the retrieved documents, the tool definitions, the conversation history, the memory of past interactions — so the agent can act correctly within a limited space. It is not the same as writing a clever prompt; it’s the broader practice of curating the entire payload an agent reasons over, treating that payload as a scarce, carefully allocated resource rather than a bottomless container.

The Old Way

Before context engineering existed as its own recognized discipline:

Early large language model applications treated the prompt as the entire problem, stuffing instructions, background, and data into a single undifferentiated block of text.
There was no systematic way to decide what actually belonged in an agent’s working context versus what was simply available and got included out of convenience.
Teams discovered context problems — irrelevant information crowding out what mattered, missing details causing wrong answers — only after something already broke in production.

Treating every input as equally worth including is exactly the instinct an expedition guide would never allow, and it’s exactly what context engineering exists to correct.

What’s Changing (and Why AI Is the Reason)

Teams building with AI agents increasingly treat context as an engineered artifact with its own review process, rather than an afterthought bolted onto a prompt.
This is a distinct discipline from wording the instruction itself, which is what this content library’s dedicated prompt engineering series covers — that series is about the note taped to the gear, while this one is about deciding what gear goes in the bag at all.
As agents take on longer, more autonomous, multi-step tasks, the cost of a poorly packed context — missing a tool definition, drowning in irrelevant retrieved documents — compounds across every step, making deliberate context curation a reliability requirement rather than a nice-to-have.

The Metaphor, Fully Extended

The Expedition Backpack	Context Engineering Concept
The guide laying out every candidate item before deciding	The engineer reviewing every candidate input before including it
The pack’s finite carrying capacity	The agent’s finite context window
Choosing gear suited to this specific route and weather	Selecting data, tools, and instructions suited to this specific task
A well-packed bag getting the traveler through the trip safely	Well-engineered context getting the agent to a correct, reliable outcome

For Beginners: What to Actually Do

Before building an agent, write down every category of information it might need — instructions, data, tools, history — before deciding what actually goes in.
Practice distinguishing “available” from “necessary”: just because a piece of information exists doesn’t mean it belongs in the agent’s context.
Get comfortable asking, for any given task, “what does the agent actually need to see to do this correctly?” before writing a single line of prompt.

For Practitioners and Leaders: The Deeper Layer

Treat context curation as a first-class engineering discipline with its own review, testing, and versioning, not a byproduct of prompt iteration.
Separate the concerns explicitly: prompt wording lives in this content library’s dedicated prompt engineering series, while what data, tools, and memory populate that prompt’s surrounding context is this series’ focus.
Build context curation into the agent’s operating loop itself, so what gets packed adapts task by task rather than staying fixed for every request.

Quick Recap

Context engineering means deliberately deciding what an agent sees, not just how an instruction is worded.
It treats an agent’s context as a scarce, finite resource, much like a backpack’s carrying capacity.
This differs from prompt engineering, which focuses on wording rather than curation.
As agents take on longer, more autonomous tasks, deliberate context curation becomes essential to reliability.

Where This Fits in the Series

This opening article introduces context engineering itself and the backpack metaphor this whole series builds on. Article 2 goes one level deeper into the constraint that makes packing necessary in the first place: the context window’s finite carrying capacity.