Claude Agent Pipeline
harinathsrinivas/claude-agent-pipeline
Pipeline multi-agente portável para Claude Code, configuração por projeto
Ver no GitHubUm plugin de orquestração multi-agente para Claude Code que implementa um pipeline genérico e agnóstico de linguagem (planejador → orquestrador → executores → juiz → git-agent → arquiteto). Toda configuração específica do projeto vive em um único arquivo PROJECT_PROFILE.md, permitindo reutilização dos agentes em qualquer linguagem.
Instalação
npx convez-ai harinathsrinivas/claude-agent-pipeline1 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
Casos de uso
- Automatizar fluxo de desenvolvimento (planejamento, execução, review, git)
- Orquestrar múltiplos modelos Claude para diferentes etapas de codificação
- Padronizar processos de desenvolvimento em projetos Java/Spring e outras linguagens
Veredito

Ferramenta madura e bem-arquitetada para orquestração de agentes Claude em fluxos de desenvolvimento. Excelente para equipes que precisam padronizar processos multi-agente, mas requer compreensão clara do pipeline e dedicação à manutenção do PROJECT_PROFILE.md.
README
A tradução para o português deste README ainda está na fila.
claude-agent-pipeline
A Claude Code plugin that adds a generic, language-agnostic multi-agent coding pipeline
(planner → orchestrator → executors → judge → git-agent → architect) plus two helper commands.
Installs the karpathy way — entirely inside Claude Code, no shell scripts. The agents carry no
project- or language-specific content; everything that varies per project lives in one file you
fill in (.claude/PROJECT_PROFILE.md). Examples lead with Java/Spring.
Install (inside Claude Code — no scripts)
In any Claude Code session:
/plugin marketplace add harinathsrinivas/claude-agent-pipeline
/plugin install agent-pipeline@claude-agent-pipeline
/agent-pipeline:setup ← optional: applies recommended global settings (Opus / xhigh)
The 8 agents are now available globally, in every project — run /agents to confirm
(planner, orchestrator, architect, judge, executor-opus, executor-sonnet,
executor-haiku, git-agent).
…or just tell Claude
"Add the marketplace
harinathsrinivas/claude-agent-pipeline, install theagent-pipelineplugin, then run/agent-pipeline:setup."
Claude Code runs the /plugin steps and the setup for you. After installing, run /reload-plugins
(or restart) if the agents/commands don't appear immediately.
Per project
Open the project in Claude Code and run:
/agent-pipeline:init-project (optionally: /agent-pipeline:init-project a short description)
This seeds CLAUDE.md + .claude/PROJECT_PROFILE.md (never overwriting existing files), auto-detects
your build/test command and entry point, and helps you fill the TODO:s. After that you only ever
edit one file — .claude/PROJECT_PROFILE.md — and the agents read everything project-specific
from it by section number. examples-spring/ is a filled-in Spring Boot reference.
What's in the plugin
.claude-plugin/
marketplace.json # marketplace manifest (this single-plugin repo)
plugin.json # plugin manifest
agents/ # 8 generic, language-neutral agents (auto-installed globally)
commands/
setup.md # /agent-pipeline:setup — recommended global settings
init-project.md # /agent-pipeline:init-project — seed + fill a project's templates
project-template/ # the templates the init command copies into a project
CLAUDE.md
.claude/PROJECT_PROFILE.md # the ONE per-project knobs file (Java/Spring-first examples)
.claude/env.template, settings.local.starter.json
examples-spring/ # a filled-in example (Spring Boot orders-service)
AGENT_WORKFLOW_NOTES.md # effort tiers + the depth-1 orchestration rule
What each agent reads from PROJECT_PROFILE.md
| Agent | Reads |
|---|---|
| architect | §1 entry points |
| planner | §2 test command, §3 testing setup |
| executor-opus / -sonnet / -haiku | §3 fixtures + isolation hazards + real-data guards, §2 test command |
| git-agent / orchestrator | §4 PR/issue-code convention, merge/archive gates |
| judge | nothing project-specific |
Notes
- Why a
/setupcommand? A plugin can ship agents and commands automatically, but Claude Code does not let a plugin set yourmodel/effortLevel./agent-pipeline:setupapplies those (non-destructively) so "install everything" is still one or two commands. - Behavioral guidelines (karpathy): kept separate on purpose. If you want them too:
/plugin marketplace add multica-ai/andrej-karpathy-skillsthen/plugin install andrej-karpathy-skills@karpathy-skills. - Updating: push changes here; on your machine run
/plugin update agent-pipeline(or/reload-pluginsduring local testing).
Design in one line
Agents = generic + installed globally + never edited. Per-project knobs = one PROJECT_PROFILE.md
the agents read by section number. Effort tiers and the depth-1 orchestration rule are in
AGENT_WORKFLOW_NOTES.md.