Thread by @aakashg0

Context engineering is the new prompt engineering.

Here's everything you need to know:

1. Etymology

It's unclear who coined the term, but two folks have been particularly important in its rise:

Andrej Karpathy and Dex Horthy.

2. What It Is

It's all about thinking beyond the prompt - thinking about ALL of the tools you have to drive success.

5 are the most important to understand:

a. RAG
b. Memory
c. State/ History
d. Prompt Engineering
e. Structured Outputs

3. Why It Matters

While it's somewhat important for LLM product features, it's incredibly important for AI agents.

Here's why:

a. RAG

When the agent doesn’t “know” enough, you dynamically inject relevant content

b. Memory

So the agent remembers what you've told it before, you store key information from previous conversations

c. State/History

To keep track of multi-step tasks, you maintain context about what's been done and what comes next

e. Structured Outputs

In order to get consistent outputs from the agent...

You tell it exactly how to structure the output, kind of like code (eg, id + query + source_type)

4. How to Implement It

Start by mapping your agent's workflow. Identify where context gaps occur, then layer in these tools.

Build your first agent.
Experiment in small prototypes.
Layer in these techniques as needed.
Create a great evals system to "hill climb."

Hope this helps 💜
Aakash

P.S. What agents have you been building?