When to Split Agents

Not every task needs a sub-agent. Know when to spawn.

Spawn a Sub-Agent When:

1. The Task is Long-Running

If a task takes more than a few minutes:

• Web research across multiple sites
• Processing large datasets
• Code generation and testing

Your main agent stays responsive while workers handle heavy lifting.

2. The Task is Parallelizable

If you can do multiple things at once:

• Research competitor A AND competitor B
• Check email AND calendar AND news
• Test on staging AND production

Sub-agents can work in parallel.

3. The Task Needs Isolation

If you don't want results mixed with main context:

• Exploratory research (might be irrelevant)
• Testing ideas (might fail)
• Sensitive operations (separate memory)

Sub-agents have their own isolated sessions.

4. The Task is Specialized

If the task needs a different approach:

• Research vs execution
• Writing vs analysis
• Creative vs technical

Different sub-agents can have different "souls".

Don't Spawn When:

• Quick questions (just answer them)
• Simple tasks (overhead not worth it)
• Context is critical (sub-agent won't have it)
• One-off actions (direct execution is fine)

The Overhead

Spawning a sub-agent has costs:

• API calls for new session
• Time to spin up
• Loss of main context

For quick tasks, direct execution wins.

Rule of Thumb

If you'd tell a human colleague "go work on this and come back when done" — spawn a sub-agent.

If you'd just do it yourself in 30 seconds — don't spawn.