Vibe Coding
Anthropic's guide says don't block the lead agent while subagents run
The 'Let the lead agent keep working while subagents run' section reports that on coding tasks, letting the lead continue instead of waiting lowers average time to completion at similar quality, token usage, and cost. The setup is three parts: the subagent-start tool returns immediately, each result comes back in a later `user` message, and the lead gets a separate tool it can call when it wants to wait. Anthropic notes the model still often chooses to wait, and the savings come from the runs where it carries on.
Source
↳ Follow the thread