Agent Harness Develop Book

What Agent Harness Develop Book does

The book defines an agent harness and walks through a four-layer architecture and the core modules of production-grade agents. The source projects it names are Claude Code, OpenCode, OpenClaw, Hermes, DeepSeek Harness, Codex and Pi. Chapters cover the agent loop, with termination conditions, error recovery and parallel tool calls, along with context compaction, memory, multi-agent collaboration patterns, agent self-evolution and loop engineering.

A September 2026 changelog in the README lists additions based on 20 engineering articles and source reading of seven projects. These include a master comparison table of six projects, a section on four multi-agent patterns (workflow, supervisor, hierarchical and swarm), a chapter on self-evolution, and notes on the three context constraints of lost-in-the-middle, context rot and attention cost. PDF versions are provided in English and Chinese and can be viewed on GitHub, and the repository itself is mostly HTML.

Key features

• Covers seven agents including Claude Code, Codex, OpenCode, OpenClaw and Hermes
• Chapters on the agent loop, context compaction, memory and multi-agent patterns
• A chapter on agent self-evolution and a section on loop engineering
• Master comparison table of six projects across architecture, loop, memory and permissions
• English and Chinese editions, including PDF versions

When to use it

• Learning how production agent loops handle compaction and error recovery
• Comparing how different agents approach memory and permissions
• Choosing a multi-agent pattern for your own agent runtime

Who it is for: Engineers who build or study agent runtimes and want a source-based comparison of existing designs.

How it fits with Hermes Agent

Hermes is one of the seven agents whose source informs the book; it is a general harness reference rather than a Hermes-specific guide.

FAQ