Multi-Agent

Multi-agent collaboration accomplishes more complex tasks by having multiple specialized agents divide responsibilities and work in coordination.

Typical frameworks for Multi-Agent architecture include Microsoft's AutoGen, CrewAI, and LangGraph.

Multi-Agent collaboration entails more context passing, higher token consumption, and longer processing time, and whether the benefits of separation of responsibilities outweigh the coordination overhead needs to be weighed in specific scenarios

Multi-Agent architecture is better suited to engineering scenarios where subtasks can be clearly decomposed and are relatively independent, rather than being the default architecture for all tasks.

Agent Skills adopts a vertical capability scaling approach within a single agent, which is more lightweight and easier to debug than Multi-Agent approaches.

Common failure modes in multi-agent architectures include infinite loop calls between agents, amplification of erroneous information along the transmission chain (hallucination cascade effect), and deadlock issues when some agents time out.

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

绝大多数生产场景从单Agent模式起步就够了，不应过早引入Multi-Agent的复杂度

Multi-Agent has two common organizational approaches: orchestrator-workers and pipeline-style.

The course compared the differences among three architectures: B2B networks, Multi-Agent, and Skills Agent.

In fields with extremely high accuracy requirements such as finance, the coordinator's instruction log (Trace) should be auditable to allow humans to quickly locate problem nodes.

The potential risk of multi-agent architecture is that if the coordinator, as an LLM, misunderstands instructions, errors will propagate downstream, and such issues are harder to diagnose than errors from a single model.

The inputs and outputs of each hop in a multi-agent chain can be observed independently, making fault diagnosis more granular than the internal reasoning process of a single large model.

In real-world projects, hybrid solutions exist where a small number of dedicated Agents handle core modules requiring parallel decision-making, while the remaining functionality is mounted in the form of Skills

In the Skills architecture, skill switching occurs at the prompt level without launching a new inference process, making it more lightweight than Multi-Agent

The cost of a multi-agent architecture is that total token consumption increases multiplicatively, because the same text needs to be read and processed separately by multiple agents.

The architecture of strong-model orchestration combined with weak-model execution is commonly referred to as the LLM-as-Orchestrator or Multi-Agent architecture, decomposing complex tasks into two layers: planning and execution

The per-session cost of a multi-agent architecture grows linearly or even super-linearly with the number of concurrent agents.

Interview requirements for large model-related positions have risen noticeably in recent times, and Multi-Agent has become a high-frequency exam topic.

In multi-agent systems, collaboration among large language models has long relied on converting one model's output into natural-language text and feeding it as a prompt to the next model.