Multi-Agent

Hierarchical multi-agent solutions often face challenges such as high coordination overhead, high token consumption, and difficult debugging when deployed in practice.

Common architectural patterns for multi-agent systems include the Supervisor-Worker pattern, the pipeline pattern, and the debate pattern.

The core advantage of multi-agent workflows is that each Agent can use a lighter-weight and more specialized model, making the overall system's reliability and controllability superior to solutions that rely on a single all-purpose large model.

Teacher Li Jin's research direction over the past two years has been to deeply combine traditional microservices with multi-agent systems and RAG retrieval, with end-to-end implementation capability from underlying algorithms to AI applications.

The applicable boundary of multi-agent systems is usually: when a task can be clearly split into independent sub-problems, when cross-validation from multiple perspectives is needed among subtasks, or when a single Agent's context window is insufficient to hold all the information.

Prompts solve 'how to communicate with the model,' RAG solves 'how to let the model use external knowledge,' workflows solve 'how to string multiple steps together,' and multi-agent systems solve 'how to let multiple roles collaborate.'

In practice, most business scenarios can achieve similar results with a single Agent plus workflow orchestration, and with greater stability and controllability.

Transportation systems are a typical multi-agent scenario, where each driver, bus, and traffic light makes local decisions, and the overall effect depends on the superposition of those decisions.

应只在出现特定协调需求时才增加第二个智能体，大多数需要多智能体的问题实际上是需要更窄工具和更好评估的问题

For tasks with relatively linear workflows, a single-model pipeline with structured tool-calling and strict output validation is cleaner, more controllable, and less prone to hallucination than a multi-agent system.

Multi-agent architectures introduce a Trust Propagation attack surface, where content transmitted by a compromised agent may carry malicious instructions, constituting a lateral movement attack.

In multi-agent architectures, messages sent from one agent to another should be protected in the same way as user input, and each agent should only possess the permissions necessary to run its tools.

Multi-agent systems should only be considered after a single agent is working correctly, because inter-agent communication means more tokens, more drift, and harder debugging.

多智能体架构相比单一大模型对话能从不同角度施压，减少某一盲区被忽略的概率

Premortem采用多智能体架构，将质疑任务分解给多个专门化的AI角色，分别质疑市场规模、财务假设、竞争对手视角等

Directions for mitigating erroneous consensus in multi-agent systems include introducing heterogeneous model sources, designing dissenter agents, integrating external tool fact-checking, and remaining vigilant about the speed of convergence.

The core difference between workflows and multi-agent systems lies in the trade-off between determinism and autonomy.

For complex business automation scenarios, multi-agent systems are considered more suitable than a single oversized model.

Multi-agent systems adopt a modular design, so adding new features or agents only requires adding new modules without restructuring the entire system.

Multi-agent, multi-round verification workflows entail a large number of LLM calls, and computational costs cannot be ignored.