Harness Architecture

paper

world 02 / editorial interferencepaper modeframe 0000scroll 0%

issue 02 / first spread

The homepage gives a verdict before it sells a concept.

This landing page is arranged like an editorial desk: the oversized title carries recognition, the image reads like a marked-up proof wall, and the index below turns 22 chapters into reviewable engineering questions.

02

white paper / black type / one red interrupt

dark mode keeps the editorial grid, swapping paper and ink instead of inverting screenshots.

contents / no generic best practices

Enter the architecture like a magazine contents page.

1. 01

   Start with the overview

   Decide whether your agent should stay light, go heavy, or borrow one layer.

   00
2. 02

   Dissect the runtime loop

   Observe, plan, act, and verify are recovery boundaries, not just a diagram.

   00
3. 03

   Compare the four systems

   Put each decision through Codex, Claude Code, OpenClaw, and Hermes.

   00
4. 04

   Package the companion skill

   The reading ends as operating memory you can load before the next implementation.

   00

reference systems / collisions

The four systems behave like evidence pinned into the layout.

Codexcoding loop and sandbox layeringRust workspace / local + cloud agentClaude CodeIDE-first UX and skill engineTypeScript / Node developer toolOpenClawmulti-channel control plane and session routingpersonal agent control planeHermeslong-running self-improvement with restrained safety boundarieslong-running Python agent

colophon / companion skill

The last page is an operating index, not a CTA.

build-your-own-agent compresses the 22 chapters into 10 rules, an 8-dimension decision table, a Python scaffold, diagnostic scripts, and safety checks. It is an editorial manual you can load before the next build.

Open companion skill