Agentic Engineering

Chapter 4. Context Engineering

A Note for Early Release Readers

With Early Release ebooks, you get books in their earliest form—the author’s raw and unedited content as they write—so you can take advantage of these technologies long before the official release of these titles.

This will be the 4th chapter of the final book.

If you’d like to be actively involved in reviewing and commenting on this draft, please reach out to the editor at [e-mail].

The user types under 5% of what the model reads. This chapter is about the other 95%.

For a couple of years, the industry called this work “prompt engineering,” and the phrase implied a magic incantation waiting to be found. Job listings chased it. Conference talks chased it. In hindsight, the framing was a photograph of where we were.

Here’s what changed. Working with a coding agent, you aren’t shaping the initial input to a prompt. You’re operating on the working memory of a system across many turns. The user prompt, the system prompt, the AGENTS.md, the accumulated tool outputs from the last 40 steps: different memory slices, competing for a limited sliver of attention inside the model. In practice the user prompt is often less than 5% of the tokens in the window, so 95% of what you’re trying to influence is the output of your harness rather than the sentence a person typed (see Figure 4-1).

Figure 4-1. A fixed-width context ...