HeadsUpAI

Simon Willison Publishes Guide on How Coding Agents Work Under the Hood

· Updated

Simon Willison published a new chapter in his Agentic Engineering Patterns guide explaining how coding agents work. The guide defines a coding agent as an LLM wrapped in a harness — software that extends the model with a system prompt, callable tools, and a loop that replays conversation state. It covers token mechanics, chat templates, cached input tokens, tool calling, and how reasoning lets models spend more time on harder problems.

This fills a practical gap for anyone using Claude Code, Codex, or Cursor — understanding the underlying loop helps you make better decisions about prompt structure, cost management, and when to increase reasoning effort. A follow-up chapter covers subagents now shipping across seven major coding agents.

Read through the guide alongside your current coding agent workflow — the mental model of "LLM + system prompt + tools in a loop" reframes how you structure tasks and debug agent behavior.

View the full update on simonwillison.net
Simon Willison
Simon Willison
@simonw
X

New chapter for Agentic Engineering Patterns: I tried to distill key details of how coding agents work under the hood that are most useful to understand in order to use them effectively https://t.co/14ai3UP79C

55retweets
View on X

Share this update