Dive into Claude Code

What Dive into Claude Code does

The repository accompanies a paper on Claude Code version 2.1.88, covering about 1,900 TypeScript files. It walks through the agentic query loop, permission and safety layers, extensibility mechanisms, context and memory handling, subagent delegation and session persistence. Its headline claim is that AI decision logic is a small slice of the codebase, and that the rest is deterministic infrastructure such as permission gates, context management, tool routing and recovery.

Beyond the paper, the README adds a design guide for building your own agent, a section on newer signals in the agent design space, a catalog of community analyses and projects organized by design question, and a cross-system comparison of Claude Code, OpenClaw and Hermes-Agent. The architecture analysis itself describes Claude Code, not Hermes Agent, so the comparison section is the part that speaks to Hermes users.

Key features

• Source-level architecture analysis of Claude Code v2.1.88
• Documentation of seven safety layers and five compaction stages
• Design guide for building an AI agent system
• Cross-system comparison of Claude Code, OpenClaw and Hermes-Agent
• Catalog of community projects and research organized by design question
• Link to the full paper on arXiv

When to use it

• Comparing how Claude Code and Hermes Agent handle permissions, memory and subagents
• Learning how production coding agents are structured before building a custom agent
• Finding community research on coding-agent architecture

Who it is for: Agent builders and researchers who want a structured study of how a production coding agent is put together.

How it fits with Hermes Agent

It is not software for Hermes Agent. It supports Hermes users as a reference, through a comparison of Claude Code, OpenClaw and Hermes-Agent.

Note: The architecture analysis covers Claude Code v2.1.88 rather than Hermes Agent.

FAQ