Sub agents vs Agent Teams

My notes

Summary

In Claude Code, sub-agents and agent teams serve fundamentally different purposes. Sub-agents are isolated, single-task workers that run in parallel without polluting the orchestrator’s context window. Agent teams are long-running, collaborative groups that share state and respond to each other in real time.

Key Insight

The core decision rule is simple: isolated work -> sub-agents, collaborative work -> agent team.

Sub-agents:

• Each runs in its own context window, keeping the main orchestrator clean
• Fire-and-forget: give a task, they complete it, return result
• Key advantage: parallelism, you can run multiple sub-agents simultaneously for independent tasks
• Best for: parallel file processing, multi-source research, running several independent pipelines at once

Agent teams:

• Long-running, they persist across tasks, not just one job
• Can directly communicate with each other and react to peer progress in real time
• Build shared context over time (useful for iterative, dependent work)
• Typical structure: lead agent + working agents + shared task state
• Best for: complex projects where agent A’s output feeds agent B, or where real-time coordination matters

Practical decision heuristic (straight from the clip):

• 3+ independent subtasks -> sub-agents, fire in parallel in the same message block
• Tasks depend on each other or need coordination -> agent team