// Track

Multi-Agent Orchestration

From Single-Agent to Fleet

From single-agent to fleet — design orchestration layers, coordinate parallel agents, manage shared state, and build systems where AI agents hand off work to each other.

8 lessons~80 min total

Lessons are shown in recommended order. Complete them in sequence for the best experience — or jump to any lesson.

Lesson 64·10 min read

The Orchestration Mental Model: Why One Agent Isn't Enough

A single agent has a ceiling. Context fills up. Execution serializes. Specialization competes with breadth. The moment you understand why one agent is architecturally insufficient, you start building fleets.

Lesson 65·10 min read

Agent Roles and Specialization: One Agent, One Job

The quality ceiling of a multi-agent system is set by how well each agent is specialized. An agent that does two jobs does both jobs worse. Role taxonomy, system prompt design, and context pollution — the architecture of focused agents.

Lesson 66·10 min read

Task Decomposition: Breaking Work Into Agent-Sized Pieces

Decomposition is the orchestrator's primary skill. What makes a task agent-sized? How do you map dependencies? What does a handoff contract look like? The fan-out/fan-in pattern and the dependency graph that keeps your fleet coherent.

Lesson 67·10 min read

Shared State and Agent Memory: The Coordination Layer

Agents are stateless by default. They finish a run and remember nothing. Coordination across agents requires an explicit shared state layer. File-based state, lock mechanisms, event logs, and memory systems — the infrastructure that makes multi-agent systems coherent.

Lesson 68·9 min read

Inter-Agent Communication Patterns: Polling, Events, and Queues

How do agents in a fleet know when to act? Polling wastes cycles. File watchers and trigger files are free. Queue-based messaging scales. The communication pattern you choose determines your fleet's efficiency, latency, and operational complexity.

Lesson 69·10 min read

Blast Radius: Keeping Parallel Agents Safe

Parallel agents amplify speed. They also amplify mistakes. A buggy agent touching a shared repository at scale can corrupt everything. Isolation layers, worktree strategies, rollback protocols, and the 2-failure rule for fleets — keeping blast radius minimal.

Lesson 70·10 min read

Orchestration Platforms: Build vs. Buy

LangGraph, CrewAI, AutoGen — or shell scripts and file state? The build-vs-buy decision for orchestration platforms is not about capability. It is about the coordination requirements your specific workflow actually has. Most workflows do not need what platforms provide. Some do.

Lesson 71·11 min read

A Production Multi-Agent System: Architecture Review

Every concept from this track — roles, decomposition, state, communication, isolation — assembled into one production architecture. The content pipeline as a live case study: triggers, agents, state layer, failure modes, and fleet health monitoring.