Routing isn’t just handing work away. It’s choosing the right path: who has the right context, authority, and fit for this exact piece of work. Offloading whatever you don’t want to do is how you create rework.
Think about a general contractor in a half-finished house. Drywall, plumbing, electrical, and inspections all need to happen. A good contractor doesn’t grab the nearest person and say, “You handle it.” They route each job to the right trade, in the right order, with the right instructions. That’s routing. The orchestrator looks at the work and decides where it belongs. Some goes to a person, some to a subagent, some stays with the operator because it needs judgment, and some shouldn’t move yet because the input isn’t ready.
How it shows up
In AI work this happens constantly. You might ask Claude Code to update a small script but keep pricing strategy yourself. You might send transcript cleanup to a cheaper model and final synthesis to a stronger one. Routing also decides the handoff: what the next owner needs to receive, what they should return, and when they should stop and ask. If the handoff is fuzzy, the work moves but the next person is guessing. That’s why ownership matters. A routed task needs a clear owner, even a temporary one, or everyone assumes someone else has it and the human finds out only when the output is wrong.
Why you care
Good routing feels slower for about five minutes while you name the work, pick the owner, and define the handoff. Everything after that moves faster, with fewer surprises downstream. Route well and you keep AI work from becoming a pile of half-owned tasks you have to rescue later.