Agentic Engineering

Chapter 3. The Anatomy of an Agent: Model, Harness, Loop

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 3rd 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].

When I was 16, I built a web browser called XWebs for a national science competition, and I gave it an assistant named Phoebe. She sat next to the browser, slightly animated, and she could talk: press a button, say a thing, get a response. She read pages out loud. I’d wired in the ideas I most wanted from software at the time: speech recognition, text to speech, theming lifted from my Winamp obsession. What I actually wanted was a browser you could pair with to get stuff done.

Phoebe could talk. She couldn’t act. Two decades on, the models hold the conversation fine, and the conversation turns out to be the smallest part of the problem. Two agents built on the same model can behave nothing alike: two of them drew from the same weights and ran the same benchmark, Terminal-Bench, and one landed near thirtieth while the other landed in the top five. Everything I wished I could have given Phoebe, the ability to touch things safely, remember things durably, and know when she was done, is the part that decides ...