Check Whether Useful Parallelism Exists

A split pays off when the branches are independent, the task has enough value, and the coordination cost is below the gain in quality or time. Otherwise you pay for coordination without getting speed.

Anthropic reports that their research system is especially useful for breadth-first tasks with independent search directions. In the same material there's a limit: multi-agent workloads in their data used roughly 15 times more tokens than ordinary chat interactions, and tasks with a large shared context and dense dependencies fit worse. This is a vendor case study, not a universal constant, but the direction of the risk is clear.

To run your task through a fit calculator and name the grounds, use the lab.

The calculator doesn't make the architectural decision - it forces you to name the grounds. After it you need a pilot with a baseline, the same tasks, and the same acceptance gate. And the signs of a poor fit are worth knowing in advance.

• Every worker re-reads the same large context.
• One worker's result constantly changes the others' task.
• Everyone writes to the same files or one shared-state record.
• The task is cheap and errors are easy to fix with one retry.
• There's no objective outcome, so the team only discusses its own answers.

The fit is assessed. But for the pilot to prove anything, the hypothesis must be stated as a gain, not a number of agents.