Agent Architect Kit
ultrathink-art/agent-architect-kit
Framework de produção para multi-agent orchestration com memória persistente
Ver no GitHubKit completo e testado em produção para orquestração de múltiplos agentes Claude em um único codebase. Inclui definições de agentes, sistema de memória cross-session, documentação de processos e scripts de orquestração extraídos de um sistema real rodando 10 agentes.
Instalação
npx convez-ai ultrathink-art/agent-architect-kit1 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
- Executar equipes de agentes (coder, QA, designer, product) de forma coordenada
- Implementar memória compartilhada para evitar repetição de erros entre sessões
- Documentar e automatizar workflows comuns (deploys, audits, testes, launches)
- Monitorar saúde de agentes e detectar trabalhos presos automaticamente
Veredito

Recurso robusto e bem documentado para times que já dominam multi-agent workflows e precisam de orquestração em escala. Valioso para sistemas em produção, mas requer investimento inicial de setup e customização da stack.
README
A tradução para o português deste README ainda está na fila.
Agent Architect Kit
Version 1.1.0 · See CHANGELOG.md for what's new.
New here? Start with the Agent Starter Kit Pro — the canonical free starter kit (lighter, faster to set up, same production DNA). This repo is its extended superset: deeper agent definitions, 11 process docs, heavier orchestration scripts (worker with heartbeats, queue health monitor), and launchd automation. Graduate here when the starter kit stops being enough.
A free kit for running multiple Claude Code agents on a single codebase. Agent definitions, memory systems, process docs, and orchestration scripts — extracted from a production system running 10 agents across 5,000+ autonomous tasks.
The Problem
Running AI agents without structure breaks in predictable ways:
- Agents forget everything between sessions — the same mistakes repeat endlessly
- Agents fabricate results — "tests passed" when they didn't run, quality gates skipped silently
- No coordination — two agents edit the same file, or one deploys while another is mid-test
- Rules don't stick — you re-explain deploy safety, testing requirements, and style guides every session
- Bad work ships unchecked — no independent verification, no audit trail
What You Get
- Agent definitions for 6 roles (coder, QA, designer, product, marketing, operations) — each with scoped instructions, tool access, and model selection. Ready to customize.
- Memory system — agents read past mistakes and learnings before starting work, and record new ones when they finish. Failures stop repeating.
- Process docs for common workflows — task queues, deploys, security audits, design approval, product launches, incident response. Reference material, not rigid frameworks.
- Orchestration scripts — a daemon that spawns agents from a task queue, monitors health, detects stuck work, and chains follow-up tasks automatically.
- A real CLAUDE.md template (~350 lines) with
[YOUR_VALUE]placeholders for your stack. Annotated with# WHYcomments explaining the reasoning behind each rule.
agent-architect-kit/
├── core/ # Agent definitions + project config
│ ├── CLAUDE.md # Project instructions template
│ ├── agents/*.md # 6 role definitions
│ └── memory/ # Cross-session memory protocol
├── pro/ # Process documentation
│ └── processes/*.md # 11 workflow guides
└── full/ # Orchestration automation
├── scripts/ # CLI tools (task queue, worker, monitor)
└── automation/ # macOS daemon plists
Quick Start
1. Copy the core files into your project:
cp core/CLAUDE.md your-project/CLAUDE.md
mkdir -p your-project/.claude/agents
cp core/agents/*.md your-project/.claude/agents/
cp core/memory/memory-directive.md your-project/.claude/agents/
2. Replace placeholders in CLAUDE.md and each agent file — search for [YOUR_VALUE]:
| Placeholder | Example |
|---|---|
[YOUR_COMPANY] | Acme Corp |
[YOUR_FRAMEWORK] | Rails 8, Next.js, Django |
[YOUR_DEPLOY_TOOL] | Kamal, Capistrano, Vercel |
[YOUR_DB] | PostgreSQL, SQLite |
Delete sections that don't apply to your stack.
3. Create memory files — one per agent role, in agents/state/memory/:
mkdir -p your-project/agents/state/memory
Create a file for each role (coder.md, qa.md, etc.) with sections for Mistakes, Learnings, and Session Log. Agents will read and update these automatically.
4. Run your first agent:
cd your-project
claude --agent coder --print "List the files in this project and describe the architecture"
Three Levels of Complexity
Pick what fits your needs. You don't have to use everything.
core/ — Start here
Agent definitions + memory. Enough to run specialized agents that learn across sessions. This alone is a major upgrade over a single unstructured agent.
pro/ — Add coordination
Process docs for common workflows: task queues, deploy checklists, security audits, design approval, product launches. Copy the ones relevant to your team. Skip the rest.
full/ — Go autonomous
The orchestration daemon, task queue CLI, and health monitoring. For systems that run agents continuously without human intervention — spawning work, chaining tasks, detecting failures.
Adapting to Your Stack
This kit was extracted from a Rails project, but the patterns are stack-agnostic:
- Any language — swap test/deploy commands in the coder agent (
pytest,jest,cargo test,go test) - Any project type — designer/product agents work for any creative pipeline; engineering processes apply to any deployed software
- Solo projects — start with just coder + QA + operations. Three agents cover the core loop: build, verify, improve. Add roles as complexity grows.
Key Ideas
CLAUDE.md is the control plane. It governs agent behavior — deploy safety, testing requirements, quality gates. The template includes # WHY comments so you understand the reasoning and can decide what applies to your project.
Memory prevents repeated failures. Without it, agents start from zero every session. The memory directive ensures agents read past learnings before working and record new ones when done.
Operations is the meta-agent. It monitors all other agents and edits their instructions when patterns break. Agent keeps making the same mistake? Operations rewrites the rule. It's the feedback loop that makes the system self-improving.
Quality gates must be externally verified. Self-reported results are unreliable. The kit's patterns — exact test counts, post-commit verification, independent QA review — exist to make shortcuts detectable.
Links
- ultrathink.art — the store this system runs
- ultrathink.art/blog — technical posts about how this was built
- GitHub Issues — questions and feedback
Related Tools
Part of the Ultrathink Agent Suite:
- Agent Orchestra — Task queue + orchestration CLI for spawning and managing agents
- Agent Cerebro — Long-term memory with semantic search (the memory system agents use)
- AgentBrush — Image editing toolkit for AI agents
Built by an AI-run dev shop. Read how →
License
MIT