🔴 Agents & Cowork

🔴 What It Is

One subagent is an instance of Claude with its own context window, launched by the main agent for a specific task. Several run in parallel, each one returns only the conclusion — without flooding the main context. You create and configure subagents with /agents.

💡 Why learn

Full audits, broad research, and refactors split across parallel agents cut total time. And since each subagent has its own context, they can scan many files without using yours—you get the conclusion, not a dump of everything they read.

🔑 Prompt Formula

Use subagentes em paralelo para [objetivo].
Divida:
1. Subagente A: [papel específico]
2. Subagente B: [papel específico]
3. Subagente C: [papel específico]
Regras: trabalho independente, não editem ainda,
        consolidem os achados no final.

Claude no creates subagents on its own for most tasks — you just ask. Patterns like DeepClaudex package this orchestration (fan-out → verification → synthesis) into a repeatable workflow.

🔴 What It Is

One step beyond subagents: a team agents with fixed roles that communicate. The backend notifies the frontend when it’s done; the frontend sends it to QA; QA returns bugs for correction. There’s also the analytical orchestrator (AOA) — an agent that coordinates the others.

💡 Why learn

For projects with dependent stages (API → UI → tests), the team delivers a realistic software workflow, with handoff and review—not just disconnected pieces. It’s the leap from “an agent that helps” to “a team that delivers.”

🔑 Full-Stack Team Prompt

Crie um time de 3 teammates usando Sonnet:
• Backend dev  → REST API, modelo de dados, endpoints
• Frontend dev → espera o backend, então cria o React que consome a API
• QA agent     → testa o app pronto, acha bugs, gera relatório pass/fail

Handoff: backend → avisa frontend → envia ao QA.
Entregáveis: app rodando + relatório de testes + doc de como rodar.

🔴 What It Is

O Cowork is Claude’s mode focused on run tasks and automations of the day (not programming): organize files, analyze emails, generate reports, connect apps. Think of it as a digital coworker the person you delegate work to — and pay in tokens, not a salary.

💡 Why learn

Knowing which tool to use avoids frustration: not everything is "code." The dispatch from Cowork launches tasks that run in the background, and the projects store reusable context and global instructions.

🔑 Which to Use and When

🔴 What It Is

Point Claude Code to your Obsidian vault (your Markdown notes) and use them as a second brain: it reads, connects, and writes to your notes. A CLAUDE.md the foundation defines the structure, and the folder variants/ stores system versions for different profiles (freelancer, student, dev…).

💡 Why learn

Your notes become live context: Claude answers based on what you knows, not just from its training. It’s the “RAG over your knowledge” approach that Karpathy popularized—searchable external memory instead of dumping everything into the prompt.

🔑 Typical Structure

vault/
├── CLAUDE.md          # regras, contexto, como organizar as notas
├── projetos/          # notas por projeto
├── decisoes/          # log de decisões
└── variants/          # versões do sistema p/ cada perfil
    ├── freelancer/
    ├── developer/
    └── student/

Pair it with Track 5 (MCP): an embeddings/search server makes note retrieval more accurate in large vaults.

🔴 What It Is

The same building blocks (subagents + skills + memory) assembled as vertical assistants: a personal assistant (PA AI), a local assistant that runs on your machine (ClaudeClaw), an executive, a legal assistant, and even a team of 5 accounting agents.

💡 Why learn

Shows that an "agent" isn't just for code. By changing the CLAUDE.md, skills, and tools (MCP), the same Claude Code becomes the engine of a niche product — the leap from study to real-world application.

🔑 Community Examples

🔴 What It Is

Close the loop: agents that run on their own. A AutoResearch loop researches a topic, verifies and synthesizes repeatedly until everything is covered; the Trigger.dev schedules and launches agents based on events/cron; “agentic” is the mindset of chaining these steps into continuous automation.

💡 Why learn

It’s what turns Claude into a team member who works while you sleep: scheduled reviews, recurring research, automatic reports. Pair it with Routines (track 5) to trigger tasks on a schedule.

🔑 Loop-Until-Covered Pattern