Engineering insights, technical deep-dives, and perspectives on building software.
One agent is useful. Multiple agents working in parallel on the same codebase is an unsolved engineering problem. We performed a source-level analysis of seven open-source projects — OrbitDock, codex-autorunner, Squad, Open-Inspect, Multica, Symphony, and AionUi — to map the architectural design space of AI agent orchestration. What emerged are four fundamental disagreements about how humans and AI should collaborate on code.
A 513,237-line TypeScript codebase. 1,902 source files. 40 tools, 101 commands, 20 service modules, 8 task execution types. We performed a complete source-level analysis of Claude Code to answer one question: what does it actually take to turn a language model into a reliable coding agent? This is not a product review — it is a technical reference grounded entirely in source code.
We analyzed Claude Code's 1,900+ file codebase — 512K lines, 90% written by Claude itself, 85K+ GitHub stars. What we found is not just an AI wrapper — it's a terminal-native IDE with its own rendering engine, multi-agent orchestration, and a multi-layered permission system.
AI coding assistants burn thousands of tokens just figuring out where they are. The context retrieval problem is the real bottleneck — and the solutions look nothing like traditional search.