Multi-Agent System
A Multi-Agent System (MAS) is a computational system composed of multiple autonomous agents, each capable of perceiving its environment, making independent decisions, and communicating and collaborating with other agents. This architecture is widely applied in artificial intelligence, distributed computing, and robotics. Through coordination and interaction among agents, a MAS can accomplish complex tasks that a single agent cannot complete alone. Key characteristics include agent autonomy, distributed system structure, and dynamic interaction mechanisms among agents.
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.
Multi-Agent architecture fundamentally avoids the attention dilution problem of a single instance by dividing tasks so that each Agent only processes context within its own area of responsibility.
Timeline (last 90 days)
In multi-Agent scenarios, an increase in the number of Agents makes tokens grow combinatorially, and because requests are highly asynchronous and vary in length, GPU utilization is often lower than in batch inference, further raising the actual compute cost per task.
In the short term, the compute cost of multi-Agent systems will limit their large-scale commercialization, and teams that can be the first to solve the inference efficiency problem will gain a significant cost advantage.
Inference Economics: most APIs currently price by token, but multi-Agent systems generate large amounts of "internal tokens" that users cannot see (Agent-to-Agent communication and self-check content), and this consumption is billed; if cost forecasts are made only based on externally visible interaction rounds, real spending will be systematically underestimated.
Cost forecasts based on per-inference pricing may be severely too low in multi-Agent scenarios; product teams estimating gross margin in the traditional way may find after scaling that actual compute spending far exceeds expectations.
Multi-Agent systems usually require an arbitration mechanism or verification layer, which adds extra compute overhead, and the more complex the system, the higher the proportion of compute used for coordination and verification.
Multi-Agent Systems are not a new concept of the LLM era and have decades of research accumulation in classical AI and distributed systems.
In multi-agent collaborative systems, each agent can only see the output handed to it by the previous agent and cannot directly verify its authenticity, and erroneous or hallucinated output may be adopted and amplified by downstream agents, forming an error chain.
The coordinator role in multi-agent systems has several mainstream implementations, including rule-based task routing, LLM-based dynamic planning, and reflection-revision loops.
Compared to single models, multi-agent architectures tend to be more stable and controllable in complex multi-step workflows, but also introduce new issues such as message-passing latency, orchestrator decision quality, and error propagation.
In multi-agent systems, confining agents to specific areas of responsibility reduces the decision-making complexity of individual agents.
21 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%VerifiedMulti-Agent architecture fundamentally avoids the attention dilution problem of a single instance by dividing tasks so that each Agent only processes context within its own area of responsibility.
90%VerifiedIn multi-Agent scenarios, hallucination problems are cascadingly amplified: the erroneous output of one Agent may become the input of the next, causing errors to accumulate continuously throughout the chain.
90%VerifiedMulti-Agent Systems (MAS) originated in the field of distributed artificial intelligence.
90%VerifiedMulti-agent collaboration accomplishes more complex tasks by having multiple specialized agents divide responsibilities and work in coordination.
80%VerifiedMulti-Agent Systems (MAS) is a classic research direction in the field of distributed artificial intelligence, with origins traceable to the 1980s.
80%VerifiedMulti-Agent System (MAS) architecture, through multiple specialized agents each performing their own duties in parallel, can circumvent the context window limitation problem of a single large model.
75%VerifiedThere are three mainstream architectural patterns for multi-agent collaboration systems: Pipeline, Hierarchical, and Peer-to-Peer.
75%VerifiedMulti-agent systems (MAS) can be traced back to distributed problem-solving research in the 1980s.
75%VerifiedResearch from Anthropic, OpenAI, and other institutions shows that multi-agent collaboration significantly outperforms single-agent approaches on long-horizon tasks.
70%VerifiedIn a hierarchical multi-agent system, there exists an Orchestrator agent responsible for decomposing tasks and dispatching work to downstream sub-agents
65%VerifiedThe mainstream multi-agent collaboration modes in the industry include Hierarchical, Peer-to-Peer, and Sequential.
65%VerifiedIn scenarios where an Agent automatically executes multi-step tasks, a failure at one step can cause the entire workflow to fail or even trigger cascading erroneous operations.
65%VerifiedIn multi-agent parallel orchestration, task decomposition granularity that is too coarse causes individual agents to suffer context overload, while granularity that is too fine generates significant coordination overhead.
65%VerifiedMulti-Agent Systems are one of the most cutting-edge research directions in the current AI field, where Agents with different specialized Skills can collaborate with a division of labor like departments within an enterprise.
80%Unverified在多智能体架构中，Orchestrator-Worker是一种经典设计范式，将复杂任务分解为编排者和工作者两个层级的角色
90%Unverified在多Agent群聊中，成员数量过多会导致上下文混乱，降低协作效率，因为每增加一个成员会更快填满可用Token空间
85%Unverified多Agent串联形成完整工作流是当前AI应用从工具走向系统的发展方向
80%Unverified多Agent系统的速度瓶颈主要来自Agent间通信开销、任务调度复杂性和API调用并发限制三个方面
80%