Why a Spawn Is a Decision That Repeats

An agent that can hand work to a helper looks like a straightforward improvement. The main agent keeps the plan, the helpers do the reading, and the answer comes back assembled. What is easy to miss is that the decision to delegate is itself available to every helper, so the structure of the work is not one agent plus some help. It is a tree whose shape nobody chose.

Consider a parent that breaks a job into three parts and gives one to each of three helpers. If each helper also delegates into three, the run has thirteen agents working before anyone looks. Nothing in that picture is broken in isolation. Each hop is a reasonable choice, and the total is still thirteen. This is why delegation depth is worth a deliberate number rather than an inherited default.

The cost is not only money. Depth multiplies latency, because each level waits on the level below it. It multiplies the chance of a flaky subtask, because a failure deep in the tree still has to travel back up. And it multiplies the volume of partial results that a parent has to reconcile, which is where context windows quietly fill.

The useful mental model is a budget. A parent has a fixed amount of work it can afford to fan out, and depth is how that budget is spent. Without a limit, the tree grows until something else stops it, and the thing that stops it is usually a timeout or a bill rather than a decision. Naming the limit in advance turns an emergent property into a choice.