Agent Coordination

Introduction

Coordination = deciding who does what when. Either centralised (supervisor) or decentralised (peers vote). Most production systems are centralised.

Beginner analogy: orchestra conductor vs jazz band. Both work; conductor scales better.

Understanding the topic

Core concepts:

• Centralised: supervisor routes tasks (simpler, debuggable).
• Decentralised: peers gossip or vote (resilient, complex).
• Use leases or locks when agents share a resource.
• Watch for deadlocks (A waits B, B waits A).
• Trace every routing decision for debugging.

Syntax reference

Visual workflow / architecture:

             ┌──────────────┐
             │ Supervisor   │
             └───┬──────┬───┘
                 │      │
                 ▼      ▼
            ┌──────┐ ┌──────┐
            │ A    │ │ B    │
            └──────┘ └──────┘

Real-world use

CrewAI hierarchical mode, LangGraph supervisor pattern, AutoGen group chat manager.

Best practices

• Default to centralised; go decentralised only for resilience.
• Log routing decisions like any RPC.

Common mistakes

• No deadlock detection — eventual hang.

Hands-on exercise

Interview preparation — practice these questions:

• Q1. Centralised vs decentralised coordination?
• Q2. How do you detect a deadlock?
• Q3. Scenario: your supervisor is the bottleneck. Two solutions?