Multi-Agent Systems

01 · 四人接力一定比一个人更可靠吗01 · Is a four-person relay always more reliable?

研究、规划、核验与写作各交给一个 Agent，看似更专业；但每次交接都有消息、费用与失真概率，整条链还能剩多少可靠率？Research, planning, checking, and writing each get a specialist. Every handoff adds messages, cost, and distortion risk—what reliability remains end to end?

02把 Agent 从 1 增至 4，观察分工、消息、成本与链路可靠率，再注入失真消息Scale from One to Four Agents; Track Roles, Messages, Cost, Reliability, and One Distorted Handoff

高阶 · 概率连乘expert · probability product

状态 AMetric A

—

状态 BMetric B

—

状态 CMetric C

—

所有概率、成本、成功率与输出均为固定种子的教学模拟，不代表真实模型输出。把主变量拖回即可复核结论。All probabilities, costs, success rates, and outputs are fixed-seed teaching simulations, not real model output. Move the main variable back to verify the conclusion.

03把刚才的变化拆成 4 个零件Break the Change into Four Parts

Multi-Agent System

ROLES

角色分工Role specialization

每个 Agent 有明确输入、输出与责任边界。Each agent has explicit inputs, outputs, and responsibility.

MESSAGES

通信拓扑Message topology

交接数量随参与者和互审边增长。Handoffs grow with participants and review edges.

COST

总成本Total cost

推理成本与通信开销必须一起计算。Inference cost and coordination overhead must be counted together.

RELIABILITY

链路可靠率Chain reliability

必经角色与消息成功率按链路连乘。Required role and message success probabilities multiply along the chain.

04最小机制：主交互能逐项验证Minimal Mechanism: Verify Each Link in the Interaction

MECHANISM

R_end = Π R_{agent i} × Π R_{message
j}

每个 Agent 与每条必经消息都小于 100%；数量增加时，即使单点很高，总可靠率仍会下降，失真消息会继续传播。Every agent and required message is below 100%. Even strong components compound downward as the chain grows, and a distorted message can propagate further.

增加角色Add roles能力分工更细more specialization

增加交接Add handoffs消息与成本上升messages and cost rise

可靠率连乘Multiply失真沿链传播distortion propagates

05四 Agent 链中注入一条失真消息Inject One Distorted Message into a Four-Agent Chain

主动制造失败Create a failure

边界实验Boundary experiment

高质量单点不能消除交接风险。实验从 Researcher→Checker 注入错误事实，并标红所有下游受影响角色。Strong individual agents do not remove handoff risk. The experiment injects a false fact from Researcher to Checker and marks every affected downstream role.

尚未执行。先在主交互中观察正常机制。Not run yet. Observe the normal mechanism in the main interaction first.

✗ Agent 越多，可靠率必然越高。More agents always improve reliability.→ ✓ 角色与消息都是新的失效点，要看端到端连乘。Roles and messages add failure points; inspect the end-to-end product.

✗ 只算模型调用费就得到总成本。Model calls equal total cost.→ ✓ 协调、复核与消息传递也消耗资源。Coordination, review, and handoffs consume resources too.

✗ 互审能自动消除失真。Peer review automatically removes distortion.→ ✓ 若共享错误前提，失真会沿消息链扩散。A shared false premise can propagate through the message chain.

06 · 一句话带走06 · One line to keep

多 Agent 分工同时增加角色与消息；总成本随协调上升，端到端可靠率按必经链路连乘，失真还会传播到下游。Multi-agent specialization adds both roles and messages. Coordination raises cost, end-to-end reliability multiplies across required links, and distortion can propagate downstream.