Multi-agent system
Also: 多代理 · 多智能體 · 多 Agent 協作 · multi-agent system
Several agents with divided responsibilities completing one task — collection, analysis and reporting split apart, joined by messages and shared state.
When you will meet it
This is the stop most distorted by marketing language. The honest ledger: it buys parallelism (five sources at once) and isolation (dirty data stays separated), and it pays in coordination failure modes (results nobody collects, two agents working from contradictory assumptions) and multiplied cost. Without that trade in view you will split single-agent work across five agents and pay five times for a worse result.
An analogy
Like a company versus a generalist. A company handles many things at once and each person remembers only their own desk — but it also meets, feuds and fumbles hand-offs; a solo worker has no such phrase as communication overhead. Five people are not automatically five times faster; the talking itself burns time.
Minimal example
單代理：一個上下文裡 收集 → 分析 → 寫報告
風險：上下文越滾越長，收集的原始雜訊污染分析
多代理（示意，本站投研流水線的分工）：
收集代理 ×N（並行）→ 產出結構化資料
↓
分析代理 → 只吃結構化資料，看不到收集過程的雜訊
↓
報告代理 → 只吃分析結論
交接處用檔案或資料庫，不用「口頭轉述」。Look at the seams: most multi-agent incidents happen not inside an agent but at the hand-offs — mismatched formats, unstated assumptions, results nobody accepts. Design the hand-off surfaces first, then decide whether to split at all.
What people get wrong
- Assuming more agents means smarter. Extra agents buy parallelism and isolation, not intelligence; when a task cannot be parallelised and needs no isolation, more agents just add cost and failure surface.
- Coordinating by letting the agents chat freely. Without explicit task assignment, result formats and acceptance conditions, free discussion decays into mutual deference or contradiction. Coordination needs orchestrator rules, not conversation.