Agent Team Pattern

A reusable architecture for coordinating multiple specialized AI agents within a single system. Extracted from oh-my-hermes.

The Pattern

Human (strategic decisions)
    │
    ▼
Orchestrator Agent (routes, monitors, reports)
    │
    ├── Specialist Agent 1 (specific domain)
    ├── Specialist Agent 2 (specific domain)
    ├── Specialist Agent 3 (specific domain)
    └── Specialist Agent N
    │
    ▼
Coordination Layer (kanban / shared state)

Core Principles

1. Specialization Over Generality

Each agent does ONE thing well. Don't make one agent do everything. The PM doesn't code. The Security agent doesn't deploy. Clear boundaries = better output.

2. Coordination Through Shared State

Agents don't communicate directly. They read/write to a shared coordination layer (kanban, task board, shared memory). This decouples agents — any agent can be swapped without breaking others.

3. Orchestrator as Router

The orchestrator (CTO in Oh My Hermes) doesn't DO the work. It: - Monitors the coordination layer - Routes tasks to the right specialist - Reports status to the human - Handles escalation

4. Human at Decision Points Only

The system runs autonomously. Human is only consulted for: - Strategic decisions (direction, priority) - Approval gates (ship/no-ship) - Exception handling (when automated flow breaks)

5. Skills Per Agent

Each specialist agent loads only the skills relevant to their role. This keeps context lean and output focused.

Application to Marketing Operations

Agent Roles

Agent	Role	Kanban Ownership
Strategist	Audience research, campaign planning, messaging	Brief
Copywriter	Headlines, body copy, CTAs, scripts	Creative
Designer	Visual direction, asset production, brand consistency	Production
Deployer	Social media, EDM, print, platform management	Deploy
Analyst	Performance tracking, A/B testing, reporting	Measure

Campaign Kanban

Brief → Strategy → Creative → Production → Review → Deploy → Measure

Human Approval Points

• Strategy review (before creative begins)
• Creative direction selection (before production)
• Deployment approval (before going live)
• Performance review (continue/pivot/stop)

Comparison with Single-Agent Approach

Dimension	Single Agent	Agent Team
Context window	Everything loaded	Only relevant skills
Output quality	Generalist	Specialist
Coordination	Implicit (same context)	Explicit (shared state)
Failure isolation	One failure = all fails	One agent fails, others continue
Human oversight	Every decision	Only strategic decisions

Related

• oh-my-hermes — Reference implementation (software dev)
• agent-knowledge-architecture — Brain vs Agent separation
• gbrain — Knowledge brain integration
• claude-design-workflow-patterns — Design workflow patterns