Skip to content
Harness Architecture
Diagram of four agent systems orbiting the Observe / Plan / Act / Verify loop

Harness Architecture

Engineering notes from four real agent harnesses: what to borrow, what to avoid, and why.
Start reading · 01 Overview System profiles Interview index

What this book covers

Section titled “What this book covers”

The book is organized by agent engineering module. Each chapter answers 5 questions:

  1. What problem does this module solve inside an agent system?
  2. How does Codex, Claude Code, OpenClaw, and Hermes each solve it?
  3. What do they agree on, and why do the differences exist?
  4. If you build your own agent, what to borrow and what to avoid?
  5. Can one architecture diagram explain the runtime?

Four reference systems

Section titled “Four reference systems”
  • Codex (OpenAI): local and cloud coding agent. Use it to study the closed-loop coding cycle and three-OS sandbox.
  • Claude Code (Anthropic): agentic coding tool. Use it to study developer UX and the skill engine.
  • OpenClaw: multi-channel personal agent control plane. Use it to study session routing and sandbox boundaries.
  • Hermes (Nous Research): long-running, self-improving agent. Use it to study engineering restraint, memory, and out-of-process security.
Observe Plan Act Verify
Every agent runs the same minimal loop: Observe → Plan → Act → Verify

How to read

Section titled “How to read”
  • Building an agent from scratch: start at 01 · Overview and read in order.
  • Copying a specific subsystem: jump to the chapter. Each §6 “Build Recipe” lists minimal, advanced, and avoid options.
  • Researching one system: see System Profiles. One page per system, with five things worth copying.
  • Each chapter’s §9 points to specific files and line numbers under REF/.

22 chapters total: an overview, plus one chapter each on agent loop, context, tool system, verifier, file edit, shell, git, code review, subagents, session, permissions, sandbox, multi-channel entry, observability/cost, memory, skills, cron, self-improvement, security, todo-list progress, and execution-state surfaces.

Companion Skill: load the book into Claude or Cursor

Section titled “Companion Skill: load the book into Claude or Cursor”

For implementation work, the book is packaged as a working index. The build-your-own-agent skill compresses the source-backed conclusions into a Claude Code / Cursor skill you can load with one command:

  • 10 iron laws (every reference system obeys them) and an 8-axis spectrum for design decisions;
  • 12-file Python scaffold ready to reuse;
  • 9 reference docs: build, diagnose, selection, refactor, security, production deploy, interview prep, cross-skill;
  • A lint script that statically checks your scaffold against the 10 laws. Drop it into CI.

See Companion Skill for install and one-line invocation patterns.