Video 21.1: Subagents vs Agent Teams

Course: Claude Code - Parallel Agent Development | Section: 21. Subagents for Parallel Work | Length: 4 minutes | Presenter: Daniel Treasure

Opening Hook

You've probably heard people talk about "agents" doing multiple things at once, but in Claude Code, there are actually two very different patterns: subagents and agent teams. Each one solves a different problem, and picking the wrong one for your task can either waste tokens or make your workflow unnecessarily complex. In this video, we'll cut through the confusion and show you exactly when to use each.

Key Talking Points

What to say:

Subagents: The Lightweight Option - Subagents are lightweight parallel workers that live within your current session - They're designed for independent, focused tasks that don't need to talk to each other - Think of them as temporary helpers you spawn for specific work, then collect their results - They have access to tools you give them, but they operate in isolation - Results come back summarized to save tokens - Much lower cognitive overhead — you're in control the whole time

Agent Teams: The Heavyweight Alternative - Agent teams are separate, peer-to-peer agents with their own sessions - They communicate with each other, share context, and make decisions as a group - Each agent has full context of what other agents are doing (task board visibility) - Great for complex workflows where agents need to negotiate or collaborate - Much higher token cost because each agent maintains full context

Key Decision: Are They Working Together or Just in Parallel? - Subagents: "Run these 3 independent analyses, then bring me the results" - Agent teams: "Coordinate as a group to architect this system, debate tradeoffs, delegate tasks" - If agents don't need to know what the others are doing → subagents - If agents need to communicate, coordinate, or negotiate → agent teams

Cost Implications (Real Talk) - Subagents: Results are summarized back to main agent. Less token churn. - Agent teams: Each agent has full session context. Scales heavily with team size. - Example: Analyzing 5 codebases - Subagents: main agent + 5 parallel Haiku searches = fast, efficient - Agent teams: 5 full agents each with complete context = expensive, but more collaborative

What to show on screen:

Open a comparison document or slide showing the two patterns side-by-side
Subagents diagram: Single session with main agent spawning temporary workers, collecting results
Agent teams diagram: Multiple sessions, agents connected on task board, communicating bidirectionally
Timeline comparison: Show parallel execution of subagents finishing in ~10 seconds vs. agent team initializing and coordinating
Token meter visualization (if available): Subagent summarization vs. full context per team member
Real example: Show the Claude Code UI for both—where you'd trigger each

Demo Plan

Setup (30 seconds): - Open Claude Code - Have a sample project ready (the same one from previous videos, or a fresh codebase) - Show the /agents command available in the prompt

Scenario 1: Subagent Use Case (1 minute) 1. Explain: "I want to analyze this project from 3 angles: code quality, security, and performance." 2. Show how you would spawn 3 subagents, each focused on one angle 3. Mention: "These agents don't need to talk to each other. I just need their independent assessments." 4. Show how results come back and you synthesize them manually

Scenario 2: Agent Team Use Case (1 minute) 1. Explain: "Now imagine these 3 agents need to design something together—architecture for a new microservice." 2. Show that agent teams have a shared task board 3. Mention: "Agent A proposes something, Agent B refines it, Agent C challenges it. They're negotiating in real-time." 4. Show the cost: more communication, more tokens, but higher-quality collaborative output

Closing Comparison (1 minute) 1. Present a simple decision tree on screen: - "Do your tasks need to talk to each other?" - Yes → Agent Teams - No → Subagents 2. Mention common scenarios for each 3. Preview: Next video will dive deep into subagents and the built-in options

Code Examples & Commands

Spawning a Subagent (Preview)

# In Claude Code, you'd use natural language or /agents command
# "Search the codebase for all database queries in parallel"
# Claude spawns 3 Explore agents automatically to split the work

Initiating an Agent Team (Preview)

# Agent teams are triggered differently—requires feature flag and more setup
# Example: natural language request with coordination needs
# "Have 3 agents debate the best approach to caching: in-memory vs. distributed vs. database"

Decision Checklist in Code

# Use this mental model:
Use Subagents when:
  - Tasks are independent
  - Results just need collection
  - Speed matters (less init overhead)
  - Token cost is a concern
  - You're the orchestrator

Use Agent Teams when:
  - Complex group coordination needed
  - Agents should communicate with each other
  - Negotiation or debate improves output
  - Token budget is flexible
  - You want emergent collaboration

Gotchas & Tips

Gotcha: "Subagents can see each other's work" - False. Subagents are isolated. One subagent cannot see another's results mid-execution. - Only the main agent collects and synthesizes results.

Gotcha: "Agent teams are always better because of full context" - Not true. More context = more tokens. For simple tasks, subagents are faster and cheaper.

Tip: Start simple, go complex only if needed - Default to subagents for parallel work - Move to agent teams only if you hit a real coordination problem

Tip: Watch the token meter - Subagents show you exactly what they found (summarized) - Agent teams can generate a lot of inter-agent chatter—monitor carefully on your first run

Tip: Use subagents for experiments - Testing 5 different approaches? Spin up 5 subagents. Much easier than team management.

Lead-out

Now that you understand the fundamental difference, let's look at the subagents you get for free in Claude Code. Claude ships with 4 built-in subagents that are instantly available—no setup needed. Each one is optimized for a specific job. In the next video, we'll meet all four, and I'll show you when to reach for each one.

Reference URLs

Prep Reading

Blog: "Why Parallel Beats Sequential" — recent Anthropic post on agent patterns
Docs: Read "Agents Overview" to understand the mental model
Review: Previous video (20.5) on agent teams for full context
Think about: What tasks in your workflow would benefit from parallel analysis?

Notes for Daniel

This video is about decision-making, not implementation. The goal is to prevent viewers from over-engineering with agent teams when subagents would work fine.

You might want to use a visual aid (whiteboard, diagram, or Figma mockup) for the architecture comparison. Seeing it beats describing it.

Tone: This should feel like a mentor saying, "Here's the shortcut—use this one most of the time, reach for that one only when you really need it."

Practice the decision tree demo. It should feel natural, not like reading from a checklist.

If possible, show a quick token meter comparison at the end (screenshots are fine). Viewers want to know cost upfront.

Quick Reference