How to Build a Coding Agent — 3-Layer Architecture

A modular coding agent should be designed with three distinct layers: (1) a provider/model layer that unifies different LLM providers and streams events transparently, (2) an agent harness layer that manages the stateless agent loop (which processes messages, executes tool calls, and returns responses) and the stateful session management, and (3) an application layer containing the terminal user interface and the actual agent with its tools, skills, and system prompts. This separation ensures the harness remains reusable and provider-agnostic while allowing different applications to consume the same event streams.