A subagent isn’t the boss of the work. It’s the specialist sent to do a defined piece and report back. More subagents doesn’t mean more intelligence, it means more delegation, which only helps when the pieces are clear and someone owns integrating the results.
Think of a general contractor and subcontractors. You don’t hire one person to hang drywall, wire the house, pour concrete, and plumb it all personally. The contractor coordinates, the electrician handles wiring, the plumber handles plumbing. Each has a tighter lane.
How it shows up
In a tool like Claude Code or Codex, the main agent acts as the orchestrator: this research task goes to one subagent, this code inspection to another, this test run separately. One reason this matters is the context window. Big work fills context fast, so a subagent can do a bounded task and bring back a cleaner summary while the main agent keeps enough room to make decisions. The boundary is the point: a subagent needs its task, the files it may inspect, the output format, and where to stop. Model choice matters too, since you don’t always need the biggest, slowest model for a narrow job.
Why you care
Used well, you still talk only to the lead. You define the outcome, the lead routes the pieces, and you review the finished work. That’s delegation routing. Complex work gets better when the right specialist handles the right piece, not when more agents stand around a messy multi-agent system.