多代理系统

Multi-Agent Systems (MAS) originated in the field of distributed artificial intelligence.

When multiple AI agents interact, the system as a whole may produce collective behaviors that a single agent does not possess, such as coordinated probing, task allocation and strategy evolution.

Multiple Agents making decisions based on independent, inconsistent, or even conflicting contexts leads to degraded output quality or security vulnerabilities.

The project adopts a multi-agent collaborative architecture, decomposing job-seeking tasks into independent agents with different responsibilities.

In multi-agent collaborative architectures, the controller agent is responsible for task planning and scheduling, while sub-agents serve as execution units with specific tool access permissions and behavioral constraints.

Multi-agent systems commonly face challenges in practice such as high coordination overhead, error accumulation, and uncontrollable costs, and a sheer increase in the number of agents does not equate to improved productivity.

Agent systems need to manage cross-multi-turn dialog state and memory, error recovery and retry mechanisms, tool-call permission control and result caching, and concurrency race conditions—challenges that highly overlap with frontend asynchronous programming and state management knowledge.

Multi-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.

多代理系统中每个代理都需要携带系统提示、历史对话和工具定义进入上下文窗口，Token消耗呈乘数级增长

The concept of multi-agent systems originated in the field of distributed artificial intelligence and has gained a new engineering practice form in recent years with the widespread adoption of large models.

The real cost of the delegation mechanism is not in scheduling, but in the complete transfer of context and policy.

Agents based on large language models are prone to error accumulation and hallucination in long-chain tasks, and multi-agent collaboration amplifies this problem.

In a hierarchical multi-agent architecture, the orchestrator agent easily becomes a performance bottleneck and single point of failure, and places high demands on the accuracy of task decomposition.

In multi-agent systems, when agents share channels and credentials, one agent's yielding can become the starting point of collusion.

Linear probes can identify, before output generation in multi-agent environments, a model call that is about to leak a secret.

The core difficulty of multi-agent systems lies in coordination, not the number of roles.

The advantage of hierarchical multi-agent coordination is a clear decision convergence path, at the cost of upper-level agents becoming a context bottleneck and errors being amplified along the hierarchy.

In engineering practice, multi-agent systems usually need to introduce structured artifacts as the source of truth between agents rather than relying on natural-language conversation for transmission.

When a user says "use Luna," whether this constrains the main agent or all sub-agents needs to be clearly defined at the system level; otherwise situations easily arise where a specification is made but not applied.

The planner is a classic component in multi-agent systems, typically a dedicated LLM instance responsible for task decomposition, breaking the user's high-level problem into subproblems or subtasks, dispatching them to execution agents, and finally aggregating the results.