Multi-Agent Planning

Introduction

When multiple agents share a goal, planning must consider who does what. Allocation, dependencies and merging become first-class problems.

Beginner analogy: a project manager assigns tickets to engineers — not random, but matched to skills and dependencies.

Understanding the topic

Core concepts:

• Plan = DAG of tasks with assignees.
• Use roles + capabilities to assign tasks.
• Mark dependencies explicitly.
• Plan critic checks completeness and conflicts.
• Re-plan when an assignee fails.

Syntax reference

Visual workflow / architecture:

Goal
 │
 ▼
DAG:
   t1 (planner)
   t2 (researcher) ◄── depends t1
   t3 (coder)      ◄── depends t2
   t4 (critic)     ◄── depends t3

Real-world use

MetaGPT plans engineering work as a DAG; CrewAI process modes encode dependencies; LangGraph supports branching workflows.

Best practices

• Represent the plan as a DAG, not a list.
• Assign roles based on capabilities, not name.

Common mistakes

• Hidden dependencies → race conditions and partial outputs.

Hands-on exercise

Interview preparation — practice these questions:

• Q1. Why DAG over list for multi-agent plans?
• Q2. How do you handle a missing assignee?
• Q3. Scenario: two agents both produce the same artefact. Resolution?