Hadley Wickham demystifies the architecture of modern AI coding agents, such as Claude Code, Cursor, and Codex. Instead of "magical" intelligence, he describes them as harness systems that operate through the orchestration of a set of basic tools.
What Happened
In his article, Hadley Wickham argues that modern coding agents are not monolithic models, but control systems equipped with six key tools: reading, writing, editing, listing files, searching, and executing shell commands. Particular attention is paid to the targeted edit mechanism, which allows for making changes to specific code fragments instead of rewriting entire files.
Context
AI tool development has traditionally focused on improving the code generation capabilities of LLMs. However, the current architecture of agents shifts the focus toward optimizing model interaction with the environment through efficient tool-use and the design of reliable harness systems.
Why It Matters for the Industry
For the industry, this signifies a transition from creating simple "wrappers" around models to designing complex control systems. A critical aspect becomes the optimization of interaction mechanisms: using targeted edits directly increases economic efficiency by reducing token consumption and improves safety by minimizing the risk of introducing errors into irrelevant parts of the code.
Why It Matters for Users
Understanding that an agent is a set of tools rather than a "black box" allows developers to take a more conscious approach to creating their own agentic workflows. This shifts the focus from endless prompt engineering to the optimization of specific tools and system control mechanisms.
What Is Not Yet Known / Limitations
No direct technical disagreements regarding the architectural composition of the tools have been identified.
Sources
Author
Look at AI, Editorial Team