One of the most useful multi-agent patterns is the research pipeline: a chain of agents where each one builds on the last.
A research agent reads files, searches the codebase, and collects facts.
A writing agent receives those facts and produces a structured draft.
A review agent checks the draft for accuracy, gaps, and quality.
I find this separation pays off at the debug stage. Each agent gets a narrow brief: the research agent focuses on gathering facts, while the writing agent handles prose only. If something goes wrong, you can re-run any single stage without repeating the others.