Context Engineering

TLDR;
Instead of throwing everything into the LLM at once, break the work into multiple steps, summarize each outcome, and save checkpoints along the way.

So at first there was LLM
Then there was prompt engineering, evals, traces, etc
Then there was a tool/ function calling.
Then the Model Context Protocol (MCP)
Now we have Context Engineering.

What is context?
All the inputs that you give to the LLM are called context. The words you type in, the pictures you upload, are all part of the context.

So what is Context Engineering?
If we consider LLM like an operating system, then the RAM is the context. Each LLM has a context window. For example, Claude 4 has 200K tokens, whereas Gemini 2.5 has 1M tokens as its context window.

So, context engineering is the process of effectively managing this context window to produce the outcome we desire.

So, the next question. How to manage the context?

Like RAM, use less wherever possible. Don't throw everything into the context all at once.

Remember these 2 concepts: Summarise and Checkpoint

Break down complex tasks into smaller pieces. Then, once the task is done, summarise the outcome. Then, preserve them as checkpoints.

But you may ask, what about all the thinking in between checkpoints? The LLM can think as much as it wants in this stage; we are not going to preserve those thoughts. Only the final summary will be preserved.

This will give us 2 advantages

1. Precise control over the outcome of each stage and the input for the next stage.
2. We can run multiple agents in parallel and combine the outputs. This can speed up our AI pipeline.

That's it. Thank you for your time. See ya in the next article.