Using several AI agents on one problem — splitting work in parallel, checking each other, or filling different roles like planner and reviewer.
One agent means one context window and one perspective. Multi-agent setups split the work: subagents explore parts of a codebase in parallel and report back; a fresh-context reviewer adversarially checks another agent's conclusion; specialist roles hand off along a pipeline. Independent perspectives catch errors a single self-reviewing agent tends to miss.
The costs are real too — more tokens, coordination overhead, and failure modes where agents agree confidently on the same mistake. The research question running through this index: when do extra agents genuinely beat one strong agent with better tools?