SPINE → Demos → Multi-Agent Playbook

The Multi-Agent Playbook

An Architectural Blueprint for Production-Ready Systems

Download the Full Blueprint

14-page architectural reference for multi-agent systems

View PDF (12.5 MB)

The Core Pattern: Closed-Loop Orchestration

"You do not prompt your sub-agents. You prompt your primary agent, and the primary agent prompts your sub-agents. They respond back to the primary agent, which synthesizes the information for you."

The General Contractor Model

  • • Orchestrator holds the "master blueprint"
  • • Sub-Agents are specialized trades
  • • Sub-Agents report to contractor, not user
  • • Direct user communication forbidden

Minimum Viable Stack

  • • Claude Opus 4.5 - Orchestrator + Sub-Agents
  • • Claude Code (CLI) - Agent harness
  • • Task Tool - Native sub-agent spawning
  • • Markdown File - Master plan

The Five Architectural Pillars

Click each pillar to explore its principle, practice, and payoff:

Pillar I: Communication

"Agents communicate through closed loops and verifiable artifacts."

The Principle

Sub-agents report exclusively to the Orchestrator. Direct user communication is forbidden.

The Practice

  • • Condense: Summaries, metrics
  • • Preserve: Files, logs, errors

The Payoff

Verifiable workflows. Task completion confirmed by artifact presence.

The Meta-Goal

"The goal is not to build the application. It is to build the system that builds the application."

We are no longer prompt engineers talking to a single model. We are system architects designing resilient, scalable, observable agentic frameworks. The agent is the new compositional unit of engineering.

SPINE Implementation

SPINE implements all five pillars of the Multi-Agent Playbook through its orchestration framework:

Blueprint Implementation Guide → SPINE Overview → All Demos →