An Illustrated Guide to AI Agents

Chapter 10. Code Agents and Code LLMs

Coding agents are some of the most potent types of AI agent in terms of the size of their likely impact. Code generation became one of the earliest applications of LLMs once they hit the scale of GPT-3 in 2020. As we saw in Chapter 3, code generation was also a major component for reasoning problems that reasoning LLMs tackle. Code is a key investment area for LLMs for two reasons. First, an enormous range of knowledge work, from software engineering to research, data analysis, and visualization, can be expressed as code. Second, and more useful for training, code can often be verified automatically: you can run it, or test it, and get a clear signal of whether it worked. That verifiability is what makes coding more tractable than open-ended tasks where there is no objective check.

This chapter walks through that world. We’ll meet the people who use and build code agents, see what makes software engineering agents distinct, build one ourselves, and finish with how the underlying code LLMs are trained.

Users and Builders of Code Agents and Large Language Models

Before we get into how code agents are built, it helps to know who they’re for, because the people who use and build them have very different needs. Figure 10-1 lays out the cast.

Figure 10-1. The target audience for code agents, ranging from end users to professional software developers ...