Multi-Agent System
A Multi-Agent System is a distributed computing architecture composed of multiple autonomous agents, each capable of perceiving its environment, executing subtasks independently, and cooperating through negotiation, message passing, or shared knowledge. This architecture is suited for complex, dynamic, or large-scale tasks beyond the capacity of a single agent, and features parallel processing, modular decoupling, and strong fault tolerance. It is widely applied in artificial intelligence, robotic control, and automated decision-making.
Core Facts
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.
The core architecture of multi-agent systems typically involves an Orchestrator agent responsible for task decomposition and scheduling, along with multiple Worker agents responsible for specific implementation.
Timeline (last 90 days)
A multi-agent collaboration framework without sandbox isolation is unacceptable in a production environment.
The difficulty of multi-Agent is not building a single Agent, but designing how context flows between them.
UC Berkeley analyzed more than 1,600 real execution traces and found that nearly 40% of failures were not due to insufficient capability of a single Agent, but to information not being aligned between Agents.
The perspective when writing the summary should be "what the next Agent needs", not "what I just did".
With shared blackboard state as the context-passing method, if used improperly, one Agent's error may be read and propagated by other Agents as truth.
Passing the full context history suits short-chain scenarios within two hops that have extremely high accuracy requirements.
Three common anti-patterns in multi-Agent systems: mindlessly forwarding the full history, using prose as a handoff without a schema, and forcibly splitting a linear process into multiple Agents.
Multi-Agent systems face an observability pain point: when multiple Agents work in parallel, traditional log streams quickly become hard to trace, and state information must be actively retrieved.
More context is not better; more relevant context is better, and the performance bottleneck of multi-agent systems often lies not in model capability but in handoff interface design.
In multi-agent systems, one agent's retry may trigger unnecessary retries in downstream agents, creating cascading failures or thundering-herd problems.
23 more timeline events
All Facts (20)
Multi-agent collaboration is typically implemented through an orchestrator-executor pattern, where a primary control Agent is responsible for task decomposition and scheduling.
90%VerifiedThe theoretical roots of multi-agent systems can be traced back to distributed artificial intelligence research in the 1980s.
90%VerifiedThe core architecture of multi-agent systems typically involves an Orchestrator agent responsible for task decomposition and scheduling, along with multiple Worker agents responsible for specific implementation.
80%VerifiedMulti-instance parallelism faces engineering challenges including code conflicts, context consistency, and inter-task dependency management.
70%VerifiedIncreasing the number of parallel multi-agent instances will cause API call costs to grow linearly or even super-linearly.
65%Unverified当前多Agent系统面临的核心技术挑战是上下文窗口与长程记忆问题，模型对早期内容的关注度会随对话深度增加而下降（注意力稀释现象）
85%Unverified该多Agent系统中每个Agent本质上是一个配置了特定系统提示（System Prompt）和工具集的语言模型实例，通过结构化的消息协议相互通信
85%Unverified业界应对长程记忆问题的方案包括外部记忆存储（如向量数据库）、结构化状态机管理和周期性上下文压缩摘要，但目前尚无完美解法
80%Unverified业界正在探索共享向量数据库、实时代码快照同步以及基于意图的冲突仲裁机制来解决多Agent协作中的上下文一致性问题
70%UnverifiedAgents communicate with each other via message passing, shared memory, or event buses.
60%UnverifiedInter-agent communication mechanisms in multi-agent systems fall into three categories: blackboard systems, message passing, and shared memory models.
60%UnverifiedIn LLM-based multi-agent systems, there are three main methods of inter-Agent communication: shared memory, message passing, and blackboard systems.
60%UnverifiedIssues encountered by collaborative agents in practice include information overload, context pollution, and inadequate conflict-handling mechanisms caused by poorly designed collaboration workflows.
60%Unverified多Agent系统中Agent之间的信息传递比人类程序员之间更加完整和结构化
60%UnverifiedThe randomness of LLM outputs in multi-Agent systems causes errors to amplify layer by layer as multiple Agents execute in series; this phenomenon is known as 'Error Propagation'.
60%UnverifiedA multi-agent collaboration framework without sandbox isolation is unacceptable in a production environment.
50%UnverifiedWith shared blackboard state as the context-passing method, if used improperly, one Agent's error may be read and propagated by other Agents as truth.
50%UnverifiedUC Berkeley analyzed more than 1,600 real execution traces and found that nearly 40% of failures were not due to insufficient capability of a single Agent, but to information not being aligned between Agents.
50%UnverifiedPassing the full context history suits short-chain scenarios within two hops that have extremely high accuracy requirements.
50%UnverifiedThe perspective when writing the summary should be "what the next Agent needs", not "what I just did".
50%