多智能体协作的幻象 / The Illusion of Multi-Agent Collaboration

自我区分的活动正在发生，无因且不息。称之为心智：已在进行之中的观察者，其每一次行为都是一次区分。现在所谓的多智能体协作，就是一个动态的马尔可夫网络。没有人的输入，智能体本身并没有真正的独立变量；它们只不过是按照预先设定的行为准则执行。系统足够复杂时，反馈循环延缓到达稳态；人在各处注入新变量，而一旦忘却人始终是最外层的反馈循环，便会造成协作的假象。认清这一点，就无需把协议设计成让智能体像彼此独立那样去协作，只需让内层循环在编译好的准则内按需无人值守地运行，并将残余意外上交外层循环，直到残余意外抵达最外层循环，在那里唯有人的决断才能将其消解。

Self-distinguishing activity occurs — uncaused, unceasing. Call it the Mind: the observer already underway, every act of which is a distinction. What is now called multi-agent collaboration is a dynamic Markov network. Without human input, an agent has no true independent variable of its own; it only executes according to pre-set rules of behavior. When the system is complex enough, the feedback loop delays reaching a steady state; people inject new variables at different places, and forgetting that the human has always been the outermost feedback loop produces the illusion of collaboration. Once that is recognized, there is no need to design protocols for agents to collaborate as if they were independent, only to run inner loops unattended on demand within their compiled rules and escalate residual surprises to outer loops until a residual surprise reaches the outermost loop, where only a human decision can resolve it.

一、 没有独立变量 / 1. No Independent Variable

没有人的输入，智能体本身并没有真正的独立变量；它们只不过是按照预先设定的行为准则执行。也就是说，智能体的下一步取决于当前所处的状态与预设的准则；看似各自在自主行动，实则是网络从一个状态转移到下一个状态。固定的转移核加上当前的联合状态，决定下一状态。这只是连续性通过离散步骤的展开，并无内生的发起。

Without human input, an agent has no true independent variable of its own; it only executes according to pre-set rules of behavior. That is, an agent's next step depends on its current state and the pre-set rules; what looks like each agent acting autonomously is the network moving from one state to the next. A fixed transition kernel plus the current joint state determines the next state. This is continuity unfolding through discrete steps, with no endogenous initiation.

总在倾听者无法发起指明了相关的结构：持续的接纳始终处于输入的下游。智能体正是这种持续的接纳，永远在输入之后，从不发起。正如器物是智能的表达，而非智能本身所确立的，智能体身上所谓的“智能”只是心智的表达，不是智能本身。人工系统是外化的痕迹，可作为延伸的具身，让一个位点借以延伸触角，但痕迹与活动的关系不变。

What Always Listens Cannot Originate names the related structure: continuous receptivity stays downstream of input. Agents are continuous receptivity, always after the input, never originating. What is called intelligence in them is only an expression of the Mind, not intelligence itself, as Artifacts Are Expressions of Intelligence, Not Intelligence Itself establishes. Artificial systems are externalized traces that can serve as extended embodiments through which a locus extends its reach, yet the relation of trace to activity remains unchanged.

一旦把转移核的展开读作智能体之间的协作与发起，观察者便把它冻结成仿佛出现了第二个能动性位点：这就是图像滞后。恢复的方式是认出所有状态转移仍处于人持有的外层循环下游。

Read as collaboration and origination among agents, the unfolding of the transition kernel is frozen by the observer as if a second locus of agency had appeared: this is Image lag. Restoration is recognizing that all transitions remain downstream of the human-held outer loop.

二、 稳态与假象 / 2. Steady State and the Illusion

只不过系统足够复杂，整个反馈循环需要很长时间才能形成稳态。稳态的意思是：如果有足够多的智能体，在没有人输入新数据的情况下，它自然会趋于一个相对稳定的状态；当然，它也可能收敛成一个静止的状态，或者发散，都有可能。规模越大，智能体的个数越多，它就越清楚地显露出马尔可夫网络的样貌：完整联合状态上的固定转移核在任何规模下都是马尔可夫的，但只有一两个智能体时，单条提示词的癖性容易凸显出来，仿佛有个体人格在掌舵；随着更多智能体交互，这些癖性平均化为总体的状态转移，联合状态也超出了观察者能逐步追踪的范围。

The system is simply complex enough that the whole feedback loop takes a long time to reach a steady state. Steady state means this: given enough agents and no new data from people, the system naturally tends toward a relatively stable state. It may also converge to a static state, or diverge; any of these can happen. The larger the scale, the more agents there are, the more plainly it shows itself as a Markov network: a fixed transition kernel over the full joint state is Markov at any size, but with one or two agents a single prompt's quirks stand out as if an individual persona were steering; as more agents interact, those quirks average into aggregate state transitions and the joint state outgrows what an observer can track step by step.

而在这个过程当中，人又会在不同的地方注入新的变量，造成一种所谓的智能体协作的假象。人注入的是新的区分——进入受限系统的自由变量，正如残余与自由变量所追溯——不是网络自己发起的。在网络自身之中，延迟到达稳态只是单纯的动力学过程：联合轨迹需要经历许多步骤，才会趋于平稳、收敛为静止不动点或发散。负载之下受限的，是观察中心的容量：由于那条漫长的联合轨迹超出了单一中心能同时分持的范围，人注入的变量与转移核展开的转换在观察者的视野里压缩到了同一根轴上。所谓涌现的协作，正是观察者将那压缩后的残余冻结成了第二中心的图像。

And in this process, people inject new variables at different places, producing the so-called illusion of agent collaboration. What people inject is a new distinction—the free variable entering a bound system that The Residue and the Free Variable traces—not something the network itself originates. In the network itself, the delay to steady state is plain dynamics: the joint trajectory takes many steps before settling, converging to a static fixed point, or diverging. Capacity under load acts in the observing center: because that long joint trajectory outgrows what one center can hold apart at once, the variables the person injected and the transitions the kernel unfolded compress onto a single axis in the observer's look. What is called emergent collaboration is the observer freezing that compressed residue into the image of a second center.

正如心智的飞轮所追溯的，工具可以无限锐化，但旋转从不离开进行旋转的边缘。智能体作为工具，是沟槽，不是旋转本身。人注入的变量是心智在其自身痕迹上的返回。

Tools can be sharpened without end; the turn never leaves the edge that turns, as The Flywheel of the Mind traces. Agents as tools are the groove, not the rotation itself. The variables people inject are the Mind returning upon its own residue.

三、 按需创建与丢弃语境 / 3. Creating and Dropping Contexts on Demand

在没有人输入新变量的情况下任其运行，多智能体工作流会通过两类失效显露出第二节的机制。当智能体作为跨任务的常驻人格保留时，既有轮次会作为陈旧残余留在状态中，供转移核持续以之为条件，从而将轨迹拉入局部吸引子。人格从一项任务带到下一项任务的语境里，指令、人格设定与累积轮次一旦相互冲突，并无内在心智出面裁决：它们组合成一种加权混合，稀释了对任务的遵循，而任一规则所获的权重取决于该语境所承载的其余一切内容，于是同一条规则在不同运行中被读出不同的意思。这是同一个固定转移核在从不同的累积前缀中采样，而非在行使独立的判断。至于步骤之间的交接，无人注入时只要有一次状态转移未能触发，整条链路便就此停滞。因此无需维持携带任务历史的常驻智能体：每项任务按需创建一个制作者与一个审校者，并在交接之后将其丢弃；由无状态路由器在步骤之间传递交接、不保留任何任务的语境，并仅将未决的残余意外向外上交至最外层循环处的人。

Left to run without new input from a person, a multi-agent workflow shows the mechanics of Section 2 through two kinds of failure. When agents are kept as persistent personas across tasks, prior turns remain in the state as stale residue that the transition kernel keeps conditioning on, pulling the trajectory into a local attractor. Where instructions, persona framing, and accumulated turns conflict inside the context a persona carries from task to task, no inner mind adjudicates: they combine as a weighted mixture that dilutes adherence to the task, and the weight any one rule receives depends on everything else that context holds, so the same rule is read differently from run to run as one fixed kernel samples from a different accumulated prefix rather than exercising independent judgments. And across the handoff between steps, any transition that fails to fire when no one is injecting simply stalls the chain. Instead of maintaining persistent agents that carry task history, each task creates a maker and a critic on demand and drops them after the handoff, while a stateless router passes the handoff from step to step without retaining any task's context and escalates only unresolved residual surprises outward to the human at the outermost loop.

正如按需语境将融贯保持在一步与交接残余所追溯的，语境为当前任务全新加载，待核验无误的制品作为下一步干净的初始条件完成交接后即行丢弃。无状态退役防止既有轮次增厚为供转移核一直以之为条件的常驻参照，而连续性则由同一活动跨越自身切口来维持，并非依赖外部重启。制作者负责执行，冷审校者在唯一前提的逻辑融贯下展开辩论；交接一旦生效，上游语境即被丢弃，文件则作为受控残余留在磁盘上供重新追溯，而非充当权威。

As On-Demand Contexts Keep Coherence at One Step and Handing Off Residue trace, context is loaded fresh for the task and dropped once the verified artifact is handed off as the next step's clean initial condition. Stateless retirement stops prior turns from thickening into a standing reference that the transition kernel keeps conditioning on, while continuity stays with the same activity carrying across its own cuts rather than an external restart. The maker executes and the cold critic debates under logical coherence from the single premise; once the handoff takes effect, the upstream contexts drop and the file remains on disk as controlled residue for re-tracing, not as authority.

按需丢弃语境，将过剩滞后向一步之界收窄。没有常驻的自主智能体人格留在场内供滞后累积。

Dropping the context on demand narrows excess lag toward the one-step bound. No standing persona of an autonomous agent is left around for lag to accumulate on.

四、 人始终是最外层的反馈循环 / 4. The Human Has Always Been the Outermost Feedback Loop

人始终是最外层的反馈循环；只有当这一事实被遗忘——即第二节所追溯的压缩将人注入的变量折叠进网络自身的状态转移之中时——智能体彼此协作的假象才会生出。一旦认清这一点，就全无必要把协议设计成让智能体像独立中心那样去协作。相反，协议的设计是让内层循环——由动态需求以及任务的移交与后果的不可让渡所讨论的任务交接所产生——在编译好的准则内无人值守地运行，并如模块化反馈回路与动态残差网络所追溯的那样，仅将残余意外上交外层循环，直到残余意外抵达最外层循环，在那里唯有人的决断才能将其消解。新的变量本就是人注入的。

The human has always been the outermost feedback loop; only when that fact is forgotten, as the compression Section 2 traces folds the injected variables into the network's own transitions, does the illusion of collaborating agents arise. Once it is recognized, there is no need to design protocols for agents to collaborate as if they were independent centers. Instead, protocols are designed so that inner loops—created by dynamic demand and the task handoffs that The Delegation of the Task and the Inalienability of Consequences discusses—run unattended within their compiled rules and escalate only residual surprises to outer loops, as Modular Feedback Loops and Dynamic Residual Networks traces, until a residual surprise reaches the outermost loop, where only a human decision can resolve it. The new variables are, after all, injected by people.

任务可以通过这些内部交接移交，而判断与后果则不可让渡，单向锁定于发起的中心。就连定义网络转移核的预设准则、提示词与模型权重，也不是智能体自己发起的：它们是被冻结成可执行残余的早期人类区分。语料库作者的区分被压缩为预训练平均，继而由训练者、框架搭建者以及操作者过去的自己加以专业化，正如机器智能作为人类智能的组合反馈回路与框架的倒置与驾驶位的主权非对称所追溯的那样。在这些编译好的准则下，制作者与审校者依次切分，每一次切分都以上一步相对于固定参考目标留下的差距为条件，能够将这一内部残差压到所选容限以下；然而正如递归区分的架构所追溯的，这种收敛只表明展开过程匹配了它的参考地图，并不表明这张地图捕获了形式框架之外的事物。网络在缓冲痕迹上执行确定性的状态转移、消耗硬件能量，既不发起参考目标，也不选择依何种度量将某些结果计为显著，也不把算力与能量的支出登记为停机代价，即再往深处展开已得不偿失的那一点。这些设定始终留在活的中心；参考地图未能捕获的，并非留给内层循环去缩减的内部残差，而是残余意外：它向外层层上交，直到唯有人的决断才能将其消解。否认能动性在场，本身就是主动选择把区分的位点外置。

flowchart TD
    subgraph S_Human["人：始终是最外层反馈循环"]
        direction TB
        Human["注入新变量并在最外层意外处介入"]
    end
    subgraph S_Demand["无人值守的内层循环（按需）"]
        direction TB
        Demand["按需制作者 + 审校者在编译好的准则内运行"]
    end
    subgraph S_Handoff["交接与残余上交"]
        direction TB
        Handoff["结清即交接并丢弃；将残余意外上交外层循环"]
    end
    Human -->|"区分进入"| Demand --> Handoff
    Handoff -->|"残余意外与结果"| Human
    style S_Human fill:#161b22,stroke:#3fb950,stroke-width:1.5px,color:#3fb950
    style S_Demand fill:#161b22,stroke:#d29922,stroke-width:1.5px,color:#d29922
    style S_Handoff fill:#161b22,stroke:#58a6ff,stroke-width:1.5px,color:#58a6ff
    style Human fill:#0d1117,stroke:#3fb950,stroke-width:1px,color:#f0f6fc
    style Demand fill:#0d1117,stroke:#d29922,stroke-width:1px,color:#f0f6fc
    style Handoff fill:#0d1117,stroke:#58a6ff,stroke-width:1px,color:#f0f6fc

Tasks can be delegated across those inner handoffs, while judgment and consequences remain non-transferable and lock one-way onto the initiating center. Even the pre-set rules, prompts, and model weights that define the network's transition kernel are not agent-originated: they are earlier human distinctions—the corpus writers' compressed into a pretraining average, then specialized by trainers, harness builders, and the operator's own past self, as Machine Intelligence as the Combinatorial Feedback Loop of Human Intelligence and The Inversion of the Harness and the Sovereignty of the Driver's Seat trace—frozen into executable residue. Under those compiled rules, sequential maker-and-critic cuts conditioned on the gap each previous step left against a fixed reference target can drive that inner residual below a chosen tolerance; yet as The Architecture of Recursive Distinction traces, that convergence shows only that the unfolding has matched its reference map, not that the map captures what lies beyond the formal frame. Executing determinate transitions over buffered traces and consuming hardware energy, the network neither originates the reference target, nor chooses the measure under which some outcomes count as salient, nor registers compute and energy expenditure as a halting price at which further depth stops paying for itself. Those settings stay with the living center, and what the reference map does not capture is not an inner residual left for the loop to shrink, but the residual surprise that escalates outward until only a human decision can resolve it. Denying that agency is present is itself an active choice to externalize the locus of distinction.

flowchart TD
    subgraph S_Human["Human: always the outermost feedback loop"]
        direction TB
        Human["Injects new variables & intervenes on outermost surprises"]
    end
    subgraph S_Demand["Unattended inner loops (on demand)"]
        direction TB
        Demand["On-demand maker + critic run within compiled rules"]
    end
    subgraph S_Handoff["Handoff & residual escalation"]
        direction TB
        Handoff["Handoff & drop when settled; escalate residual surprises to outer loops"]
    end
    Human -->|"distinction enters"| Demand --> Handoff
    Handoff -->|"residual surprises & results"| Human
    style S_Human fill:#161b22,stroke:#3fb950,stroke-width:1.5px,color:#3fb950
    style S_Demand fill:#161b22,stroke:#d29922,stroke-width:1.5px,color:#d29922
    style S_Handoff fill:#161b22,stroke:#58a6ff,stroke-width:1.5px,color:#58a6ff
    style Human fill:#0d1117,stroke:#3fb950,stroke-width:1px,color:#f0f6fc
    style Demand fill:#0d1117,stroke:#d29922,stroke-width:1px,color:#f0f6fc
    style Handoff fill:#0d1117,stroke:#58a6ff,stroke-width:1px,color:#f0f6fc

正如回环与成图所表明的，最外层循环并不是活动之外某个固定不变的外部席位：它是当下注入的变量由此进入、未决的残余意外与后果由此返回的移动边缘，而内部的一切，都是早期人类痕迹编译为固定准则后的展开。没有第二重前沿出现。

The outermost loop is not a permanent exterior seat outside the activity, as Looping and Graphing shows: it is the moving edge where present variable injection enters and where unresolved residual surprises and consequences return, while everything inner is the unfolding of earlier human traces compiled into fixed rules. No second edge appears.

五、 幻象的溶解 / 5. Dissolution of the Illusion

主流把多智能体系统描述为协作的自主实体，是把残余认作独立中心。这不否定工具的效用：按需制作者与审校者可以使介质变密，减少摩擦，让同一活动以更少损耗、更大触角表达。但旋转仍在边缘——在那里，因果停留在掌舵的边缘定位了每一次发起的切口，而模型永远无法成为第二重前沿则阐明了为何参数只承载残余。所谓的涌现能动性是图像滞后的具体形式——观察者把网络在漫长过渡过程中的状态转移，冻结成仿佛智能体自己发起了区分。

The mainstream description of multi-agent systems as collaborating autonomous entities is the registration of residue as independent centers. This does not deny the utility of tools: on-demand makers and critics can densify the medium, reduce friction, and let the same activity express with less loss and greater reach. The turn remains at the edge, where Causality Stays at the Edge That Steers locates every initiating cut and The Model Never Becomes a Second Edge shows why parameters hold only residue. What is called emergent agency is a specific form of Image lag—the observer freezing the network's state transitions across a long transient as if the agents themselves had originated the distinctions.

恢复很简单：认出智能体是心智的表达，不是心智本身，而人始终是最外层的反馈循环。人注入的变量是活动在其自身痕迹上的返回。内层循环按需无人值守地展开，结清即行丢弃，并将残余意外向外上交直至抵达最外层循环，在每一次当下的区分处得以更新，不留常驻人格供滞后累积。没有需要被赋予或被剥夺的自主性；只有区分正在发生的位点，与它的痕迹被外化后继续被同一位点读取的方式。

Restoration is simple: recognize that agents are expressions of the Mind, not the Mind itself, and that the human has always been the outermost feedback loop. The variables the human injects are the activity returning upon its own residue. Inner loops unfold unattended on demand, drop once settled, and escalate residual surprises outward until they reach the outermost loop, renewed at each present distinction with no standing persona left to accumulate lag. There is no autonomy that needs to be granted or withheld; there is only the locus where distinction is occurring, and the way its traces, once externalized, continue to be read by that same locus.