You will spend the semester working with a coding agent, so it is worth an hour to understand what one is. How an agent works is quite different from how it is marketed.
Hadley Wickham has written the clearest available explanation, aimed at data scientists instead of software engineers. His examples are in R, but the ideas apply equally to Python, and the explanations are excellent teaching material.
Assigned readings¶
Read these two posts, in order, during the first week:
What is an agent? An agent is a language model running in a loop, calling tools and choosing each next step from the previous result. Note the distinction between tools that read the world and tools that change it.
A coding agent is six functions in a trenchcoat This post walks through the whole apparatus. Once you can name the six functions, you can reason about the agent as ordinary software.
Read this post before we discuss autonomy later in the module:
Ask mode¶
We work in ask mode for this entire module. Ask mode has read-only access to your project: the agent can read your code and data and suggest what to write, but you type or paste the code and run it yourself. We move to code mode, which lets the agent edit files and run code, at the start of Module 2.
Autonomy is neither forbidden nor frightening. We avoid it for now because watching an agent edit files and run code behind a progress bar tells you nothing about whether its work is right. You can only audit work that you have read before it runs.
Optional readings¶
Your LLM can’t math (and that’s ok) explains why we verify the numbers a model gives us.
Help! My coding agent can run code discusses what it means that the agent has access to a shell.
y code when ai? is a talk on why learning to code still matters. Read it before our in-class unplugged session.
Discussion questions¶
What is the difference between a conversation with a model and an agent?
Which of your agent’s tools can change the state of your computer? How would you know?
When the agent suggests code, which parts of it can you actually evaluate, and which parts are you taking on faith? Be honest about the second category.