Multi-Agent Systems

Multi-agent collaboration is typically implemented through an orchestrator-executor pattern, where a primary control Agent is responsible for task decomposition and scheduling.

The theoretical roots of multi-agent systems can be traced back to distributed artificial intelligence research in the 1980s.

Multi-Agent Systems (MAS) is a classic research direction in the field of distributed artificial intelligence, with origins traceable to the 1980s.

The concept of Agent originates from the "intelligent agent" theory in the field of artificial intelligence, traceable as far back as multi-agent systems research in the 1990s.

Multi-agent architecture uses role separation to keep each agent's context window relatively short, theoretically reducing attention drift and enabling parallel speedup.

In multi-agent systems, a compromised agent can deliver malicious messages to other agents via the message board, creating a "lateral propagation" effect.

Internal agents that are maliciously prompt-injected or misled may execute dangerous operations while holding legitimate permissions.

The concept of multi-agent systems originated in the field of distributed AI, referring to multiple autonomous agents completing tasks through interaction, collaboration, or competition that a single agent cannot accomplish independently.

Multi-agent systems exhibit non-stationarity, as the optimal policy for each agent depends on the behavior of the other agents.

Assigning each agent a dedicated capability boundary and prompt often outperforms having a single model handle everything, a principle that is connected to the effectiveness of chain-of-thought prompting.

"Hallucination" refers to models generating content that appears authoritative but is factually incorrect, with very high confidence.

Workflow orchestration tools experience a rapid proliferation of nodes when logic branches become complex, lack version control and unit testing support, and struggle to handle complex systems requiring dynamic decision-making

The mainstream multi-agent collaboration modes in the industry include Hierarchical, Peer-to-Peer, and Sequential.

The core idea of multi-agent systems is to decompose complex problems and have multiple agents with autonomous decision-making capabilities collaboratively solve them, with each agent possessing a local perspective and specialized capabilities.

In multi-agent systems, each Agent call must carry the full context, and multi-turn reasoning continuously accumulates conversation history, meaning a seemingly simple multi-Agent task may consume tens of thousands or even hundreds of thousands of Tokens, whereas an equivalent workflow might only require a single call of a few hundred Tokens.

Models of different costs can be mixed and matched according to the importance of tasks, achieving a balance between effectiveness and investment.

A heterogeneous agent orchestration architecture can have high-capability models handle task decomposition and quality evaluation while low-cost models handle parallel execution, striking a balance between quality and cost.

Multi-agent systems suffer from coordination failure — when the local optima of sub-agents cannot combine to form a global optimum, the system's overall performance may be worse than that of a single large model.

多智能体系统将复杂任务拆解为多个子任务，由不同AI智能体并行或串行处理再汇总结果，相较单一模型能缩短处理时间并提升可靠性

Communication patterns between agents in multi-agent systems are typically divided into three categories: point-to-point direct messaging, indirect collaboration based on a shared blackboard, and a hierarchical structure with unified scheduling by an orchestrator.