🛠️How Coding Agents Actually Work, Under the Hood
TL;DR
Simon Willison's agentic-engineering guide explains a coding agent as a harness that wraps an LLM with tools exposed through invisible prompts. He walks through the building blocks and shows a usable tool loop is only a few dozen lines on top of a model API.
Simon Willison's agentic-engineering guide explains a coding agent as a harness that wraps an LLM with tools exposed through invisible prompts. He walks through the building blocks and shows a usable tool loop is only a few dozen lines on top of a model API.
Key Points
Defines a coding agent as an LLM harness extended with callable tools via hidden prompts
Covers building blocks: workspace understanding, prompt construction, tool integration, execution
Shows a basic tool loop runs in a few dozen lines over an existing LLM API
Notes a production-quality loop needs far more work around safety and recovery
Frames reasoning models as the 2025 unlock that made these agents practical
Why It Matters
Demystifying the harness lets engineers debug and build their own agents instead of treating tools like Claude Code or Codex as black boxes.
Quick Facts
Frequently Asked Questions
Why does this matter?
Demystifying the harness lets engineers debug and build their own agents instead of treating tools like Claude Code or Codex as black boxes.
What happened?
Simon Willison's agentic-engineering guide explains a coding agent as a harness that wraps an LLM with tools exposed through invisible prompts. He walks through the building blocks and shows a usable tool loop is only a few dozen lines on top of a model API.
Comments
Be the first to comment
Enjoyed this article?
Get it daily. 7am. Free. Reads in 5 minutes.
Join 1,960 builders reading daily.