Multi-agent when not to

When is a multi-agent architecture overkill compared to a single LLM with tools?

Answers use simple, clear English.