← Back to AI News
arXiv • Apr 14, 2026
Claude Code architecture research maps the design space of coding agents
A Claude Code architecture study breaks down permission modes, context compaction, MCP, skills, hooks, and agent loops.
• Coding agents are not only prompt wrappers; they are tool runtimes with safety, memory, and permission systems.
• MCP, skills, hooks, and subagents are becoming reusable architectural primitives.
• Governance around shell commands and file edits is a product requirement.
Builder take
• Any AI workstation that edits files or calls tools needs explicit permission modes and audit logs.