On this page

Every agent carries context

A subagent usually needs instructions, repository context, tool calls, and its own output. Five agents do not share one free context window. They create five streams of work, even when they begin from the same plan.

Some frameworks also add coordination messages between the parent and subagents. Those steps may be invisible in the prompt box but still contribute to usage.

Track concurrency as a first-class field

When a usage jump looks surprising, note how many agents were active, which models they used, and how long they ran. Compare the total with a sequential version of a similar job.

The question is not whether parallel work uses more capacity. It usually does. The question is whether the reduced elapsed time and broader exploration improve the outcome enough to justify it.

Keep your usage visible while you work. Get Super Notchy for $27 ↗

Set a launch threshold

Before starting parallel work, check the short and weekly windows. If either is already tight, reduce the number of agents or assign lightweight research to a cheaper or local model.

Keep activity status visible so finished or waiting agents do not continue unnoticed. Super Notchy combines allowance with working, waiting, and finished states, which makes parallel sessions easier to supervise from one place.

Sources and further reading