[How Coding Agents Work][07] One Core, Five Surfaces

How Coding Agents Work · Module 07 of 29: One Core, Five Surfaces Part I — One Concrete Turn The same core drives an interactive terminal UI, a headless CI mode, an IDE JSON-RPC server, an MCP server, and an SDK. How a submission/event protocol makes that possible, and what it teaches about building an agent as a library rather than an app. After this module you can design an agent core that survives having four front ends, and explain why the protocol boundary sits where it does. Lessons in this module: 1. The submission/event queue pattern: operations in, correlated events out 2. Protocol design rules worth stealing: a types-only crate, transport agnosticism, non-exhaustive event enums, and wire-name compatibility across versions 3. Two protocol layers, one engine: internal operations versus the product-facing thread/turn/item API, and the adapter between them 4. Interactive and headless: the terminal UI and CI mode as consumers of the same in-process client 5. Embedded: the app-server JSON-RPC surface, and Codex exposed as an MCP server Prerequisites: module 4. About the course: Built from the OpenAI Codex harness at commit 8d01cd42fa.