convez.ai
MCP ServerAutomaçãoAvançado
androidZzT

Harness Engineering — Guia Prático para Agentes de Código

androidZzT/harness-engineering-practice

Engenharia da harness: specs, skills e hooks para agentes IA de código

Ver no GitHub
InicialPluginScore · de 0 a 100

Um guia técnico prático sobre a camada de plataforma de agentes de codificação (Claude Code e Codex), cobrindo skills, hooks, configuração, permissões e esforço de raciocínio. Ensina como construir especificações e compor primitivas da harness sem modificá-la diretamente, estabelecendo limites claros entre engenharia de plataforma e engenharia de negócios.

Instalação

$npx convez-ai androidZzT/harness-engineering-practice

1 comando no terminal: detecta o formato do repositório (marketplace ou plugin avulso) e instala via Claude Code. Requer Node 18+ e o Claude Code instalado.

Métricas

Estrelas
2
Forks
0
Acompanhando
0
Saúde
42%
Último commit
há 1 mês
Versão
—
Issues abertas
0
Contribuidores
1

Casos de uso

  • Estruturar workflows complexos com agentes de código (Claude Code, Codex)
  • Configurar extensões, hooks e permissões sem reescrever a harness
  • Orquestrar múltiplos agentes com contexto e memória gerenciados
  • Controlar esforço de raciocínio e modos de permissão em agentes

Veredito

Repositório essencial para engenheiros construindo workflows sérios com agentes de IA. Oferece análise comparativa profunda entre Claude Code e Codex, estabelecendo disciplina clara: componha as primitivas da plataforma em vez de modificar a harness. Leitura obrigatória para agentic engineering em produção.

README

A tradução para o português deste README ainda está na fila.

English ∙ 한국어 ∙ 日本語 ∙ 中文

Harness Engineering

A practical guide to the platform layer of coding agents — Claude Code & Codex.


What is Harness Engineering?

When you build with a coding agent like Claude Code or Codex, your code is only half the story. The other half is the harness — the runtime around the model that manages context, memory, tool calls, sub-agents, skills, hooks, and the loop that holds it all together.

"the full environment of scaffolding, constraints, and feedback loops that surrounds the agent." — OpenAI, Harness engineering

Harness Engineering is the discipline of building on top of that layer instead of fighting it. The central claim of this series is a boundary:

The harness is the coding agent's platform layer. Your business engineering is a set of specs built on top of it — and you should not rewrite the harness itself.

Get that boundary right and a lot of agent pain disappears: context windows stop overflowing, sub-agents stop leaking across phases, and every platform upgrade is a free upgrade instead of a three-hour compatibility audit.

This is a hands-on, source-grounded series comparing how the two harnesses most engineers actually use — Claude Code (Anthropic) and Codex (OpenAI) — solve the same problems: skills, configuration directories, hooks, permissions, and reasoning effort.

Who is this for

  • Engineers building serious workflows on Claude Code, Codex, Cursor, or any agentic coding tool.
  • Anyone who has felt the urge to "just patch the agent" and wants to know when not to.
  • Readers interested in agentic engineering, context engineering, and spec-driven development (SDD) as practical disciplines rather than buzzwords.

Contents

#ChapterWhat it covers
01What "Harness" Actually MeansThe boundary between the platform layer and business engineering; why you shouldn't modify the harness
02How to Write Specs for Complex TasksMulti-agent orchestration, the orchestrator entry point, and how to organize rules / docs / skills
03Extending the HarnessSkills, configuration directories, and hooks — the two extension models of Claude Code vs. Codex
04Controlling the Agent: Permissions & EffortThe control surface — permission modes and reasoning effort — compared across Claude Code and Codex
05Surviving Long Tasks: Compaction, Memory, GoalsComing soon

The one idea

If you remember one thing:

Your job is to write the specs and compose the primitives the harness exposes — not to modify the harness itself. Restraint is the core of the engineering discipline.

Further reading

The series is grounded in primary sources. The most useful ones:

Translations

This guide is available in English, 한국어, 日本語, and 中文. The Chinese version is the original; the others are translations. Spotted a translation issue? PRs are welcome — see GLOSSARY.md for the locked terminology.

License

Licensed under CC BY 4.0 — share and adapt with attribution.

Ver o original no GitHub

Bom para

Relacionados