@PierceZhang34: Recently, Anthropic published an engineering blog post that detailed their multi-agent research system. The conclusion is quite striking: using Claude Opus 4 as the main orchestrator and Claude Sonnet 4 as sub-agents, the multi-agent system outperforms a single Claude ...

Recently, Anthropic published an engineering blog post that detailed their multi-agent research system. The conclusion is quite explosive: using Claude Opus 4 as the main orchestrator and Claude Sonnet 4 as sub-agents, the multi-agent system outperformed a single Claude Opus 4 working alone by 90.2%. Not a 10% or 20% improvement — it's 90%. This number lets us see once again the powerful capabilities of multi-agent systems. First, how they did it. The architecture is very clear: a Lead Agent decomposes problems and assigns tasks, while multiple sub-agents execute in parallel. Each sub-agent receives clear sub-goals, output format, and tool instructions — not just let loose to do whatever. For example, for a research task, the lead agent dispatches 10 sub-agents to research 75 companies, each handling 7-8 companies, running in parallel and then aggregating results. A single agent would completely fail: the context window would quickly be blown, time would be insufficient, and parallelization would be impossible. But what is the cost? The multi-agent system consumes about 15 times the tokens of a standard single agent. So Anthropic themselves said: improving model quality is more effective than doubling the token budget. But it's not just about having more agents — the right architecture plus the right model is key. Anthropic also summarized five multi-agent collaboration patterns: Generator-Verifier, Orchestrator-Subagent, Agent Teams, Message Bus, Shared State. It's not that more complex is always better; for simple tasks, Generator-Verifier is enough, and setting up a full orchestration framework would be wasteful. Multi-agent systems indeed represent a qualitative leap, but you need to make sure about three things: Is the task complex enough to require parallelization? Are the boundaries between sub-agents clear? Are you willing to bear 15x the token cost? If you can answer yes to all three, then it's worth pursuing. Otherwise, a strong model with good prompts is sufficient.