創発的役割特化を統治する安定法則: 制約密度下のエージェント企業ダイナミクス

1. はじめに

1.1 The Agentic Company

An agentic company is an organization in which a significant fraction of operational decisions are made by autonomous AI agents rather than human employees. Unlike traditional automation — which replaces specific, well-defined tasks with deterministic programs — agentic operation replaces open-ended judgment with learned policies. An agent in an agentic company does not follow a script; it observes its environment, updates its beliefs, selects actions based on expected utility, and adapts its strategy in response to outcomes. The critical distinction is that agents interact with each other. A procurement agent's decision to switch suppliers affects the logistics agent's routing optimization, which affects the finance agent's cash flow projection, which feeds back into the procurement agent's supplier evaluation. These interaction effects are not incidental — they are the defining characteristic of agentic companies and the source of both their power and their instability.

Traditional organizational theory addresses coordination through hierarchy, standardization, and formal communication channels. These mechanisms were designed for human cognitive constraints: bounded rationality (Simon, 1955), limited span of control (Urwick, 1956), and communication overhead that grows super-linearly with team size (Brooks, 1975). Agents face none of these constraints. An agent can maintain awareness of every other agent's state. It can process thousands of coordination signals per second. It can instantaneously adopt any role for which it has a trained policy. This flexibility is precisely what makes agentic companies unstable: without the natural friction of human cognition, organizational structure can change faster than it can be governed.

1.2 The Role Specialization Problem

In any multi-agent system, agents must decide what to do. When multiple task types exist — planning, execution, auditing, negotiation, research, communication — agents must allocate themselves across these roles. Centralized assignment (a scheduler assigns roles) is simple but brittle: it creates a single point of failure, cannot adapt to local conditions, and requires global knowledge that is expensive to maintain. Decentralized assignment (each agent chooses its own role based on local information) is robust and adaptive but introduces a coordination problem: how do agents converge on an efficient division of labor without oscillating, clustering, or fragmenting?

This is the role specialization problem, and it is the central challenge of agentic company design. We want agents to spontaneously differentiate into roles that collectively cover all necessary functions, allocate capacity proportionally to demand, and maintain this allocation in the presence of perturbations. We want this specialization to emerge from local agent interactions rather than global planning. And we want formal guarantees that the emergence process converges rather than oscillating or diverging.

1.3 Governance as a Control Input

The key insight of this paper is that governance is not merely an overhead imposed on the system from outside — it is a control input that shapes the dynamics of role specialization. Governance constraints reduce the action space available to each agent. An agent that is constrained by an approval requirement before changing roles will change roles less frequently. An agent that must provide evidence for its role choice will specialize in roles for which evidence is easy to generate. An agent operating under a budget constraint will gravitate toward cost-effective roles.

These effects are not side effects of governance — they are the mechanism by which governance produces order. The question is how much governance is needed. Too little governance (low constraint density D) leaves the influence propagation matrix unchecked, allowing cascading role changes that destabilize the organization. Too much governance (high D) suppresses the very flexibility that makes agentic companies valuable, freezing agents into suboptimal roles. In this paper we distinguish two thresholds: the exact contraction condition (1 − D)λ_max(A) < 1, which determines whether the linearized dynamic converges at all, and the buffered specialization envelope λ_max(A) < 1 − D, which is stricter but preserves operational headroom against model error, asynchronous updates, and transient coupling spikes.

1.4 貢献

This paper makes five contributions: 1. Model Definition. We formalize the agentic company as a time-varying graph G_t = (V, E_t, S_t, Π_t, R_t, D_t) with a five-dimensional state vector and governance constraint density (Section 2). 2. Influence Propagation Analysis. We define the influence propagation matrix A_t and characterize the spectral properties that determine system stability (Section 3). 3. Two-Level Stability Criterion. We separate the exact contraction condition (1 − D_t)λ_max(A_t) < 1 from the buffered specialization envelope λ_max(A_t) < 1 − D_t, and derive a four-regime phase diagram spanning Stagnation, Buffered Specialization, Fragile Specialization, and Cascade (Sections 5–6). 4. Multi-Layer Extension. We extend the framework to agentic companies operating within civilization-level governance layers, deriving effective constraint density そして、純粋な縮小とバッファされた操作を区別します (セクション 8)。 5. 経験的検証 111 のエージェントと 10 の役割タイプを使用した MARIA OS Planet-100 シミュレーションでバッファリングされた動作境界を検証し、96.8% の分類精度を達成しました (セクション 10)。

2. Model Definition

2.1 The Agentic Company Graph

定義 1 (Agentic Company)。 時刻 t におけるAgentic Companyは次のタプルです。 G_t = (V, E_t, S_t, Π_t, R_t, D_t) ここで: - V = {v_1, v_2, ..., v_N} は、N 個のエージェントのセット (固定母集団) です。 - E_t ⊆ V × V は、時間 t における有向影響エッジのセットです。エッジ (v_i, v_j) ∈ E_t は、エージェント v_i が時間 t におけるエージェント v_j の役割選択にゼロ以外の影響を与えていることを示します。 - S_t: V → ℝ^5 は、各エージェントに 5 次元の状態ベクトルを割り当てます。 - Π_t: V → Δ(Actions) は、各エージェントに利用可能なアクション (そのポリシー) の確率分布を割り当てます。 - R_t: V × Actions → ℝ は、エージェントとアクションのペアをスカラー報酬にマッピングする報酬関数です。 - D_t ∈ [0, 1] はガバナンス制約密度です。

2.2 状態ベクトル

各エージェント v_i には、次の 5 つのコンポーネントを持つ状態ベクトル S_t(v_i) = [F_t, K_t, H_t, L_t, C_t] があります。 |コンポーネント |記号 |説明 |範囲 | |----------|----------|---------------|----------| |財務資本 | F_t |予算、収益能力、リソース配分 | [0, ∞) | |ナレッジキャピタル | K_t |蓄積された専門知識、訓練されたスキル、モデルの品質 | [0, 1] | |ヘルスキャピタル | H_t |動作の信頼性、稼働時間、エラー率の逆 | [0, 1] | |正当性資本 | L_t |ガバナンスからの信頼スコア、監査合格率 | [0, 1] | |調整資本 | C_t |ネットワーク位置の値、影響中心性 | [0, 1] | 状態ベクトルは次のように進化します。 S_{t+1}(v_i) = f(S_t(v_i), π_t(v_i), R_t(v_i), D_t, {S_t(v_j) : (v_j, v_i) ∈ E_t}) ここで、「f」は状態遷移関数です。近隣の状態への依存 {S_t(v_j)} は、influence propagates. The dependence on D_t is the mechanism through which governance shapes dynamics.

2.3 Governance Constraint Density

Definition 2 (Governance Constraint Density). The governance constraint density at time t is: D_t = |Constraints_t| / |ActionSpace_t| where |Constraints_t| is the number of active constraints (approval gates, evidence requirements, budget limits, role-change cooldowns, audit triggers) and |ActionSpace_t| is the total number of available actions across all agents. Intuitively, D_t measures the fraction of the action space that is restricted by governance. When D_t = 0, agents face no constraints and can take any action at any time. When D_t = 1, every action requires approval and the system is fully locked. In practice, operational agentic companies operate in the range D_t ∈ [0.15, 0.70]. The constraint density is not merely a count of rules. It accounts for the interaction between constraints and the action space. Adding a constraint that restricts an action no agent would take does not increase D_t. Removing an action that was already constrained does not decrease D_t. The density is a genuine ratio that reflects the effective governance burden on agent behavior.

2.4 Role Space and Assignment

Definition 3 (Role Space). The role space ℛ = {r_1, r_2, ..., r_M} is a finite set of M functional roles. Each role r_k is characterized by a task distribution, a skill requirement vector, and a reward profile. A role assignment at time t is a function ρ_t: V → ℛ mapping each agent to a role. The aggregate role distribution is the vector ρ̂_t ∈ Δ^M where ρ̂_t(k) = |{v_i : ρ_t(v_i) = r_k}| / N is the fraction of agents assigned to role r_k. An important modeling choice is that agents can change roles at each time step, subject to governance constraints. A role-change cooldown constraint, for example, prevents an agent from switching roles more than once every τ steps. This is a governance mechanism that directly reduces the effective action space for agents who have recently switched roles, thereby increasing D_t locally.

2.5 The Markov Decision Process Formulation

各エージェントの観点から見ると、エージェント会社は部分的に観察可能なマルコフ決定プロセス (POMDP) です。時刻 t におけるエージェント v_i は、自身の状態 S_t(v_i)、その近傍の状態 {S_t(v_j) : (v_j, v_i) ∈ E_t}、および現在のガバナンス パラメーター D_t から構成されるローカル状態 o_t(v_i) を観察します。この観察に基づいて、役割「r_i(t+1)」とその役割内のアクションが選択されます。報酬 R_t(v_i, a) は、エージェントのアクション、他のエージェントのアクション (タスクの完了と調整効果による)、およびガバナンス コンテキストによって異なります。 重要な課題は、MDP が 非定常であることです。他のエージェントの役割が変わると、移行のダイナミクスも変化します。エージェント「v_i」の最適なポリシーは、他のすべてのエージェントのポリシーに依存し、エージェントは「v_i」のポリシーに依存します。これは、古典的なナッシュ均衡問題です。当社の安定性分析では、結合された MDP のシステムがどうかを検討します。converges to a fixed point (stable specialization) or cycles indefinitely (chaos).

3. 影響の伝播

3.1 The Influence Matrix

定義 4 (影響行列)。 影響伝播行列 A_t ∈ ℝ^{N×N} は次のように定義されます。 A_t[i, j] = ∂U_j(r_j | ρ_t) / ∂ρ_t(v_i) ここで、U_j(r_j | ρ_t) は、現在の役割割り当て ρ_t を前提とした現在の役割 r_j に対するエージェント v_j のユーティリティであり、偏導関数は、エージェント v_i のロール選択におけるわずかな変更がエージェント v_j のユーティリティにどのような影響を与えるかを測定します。 直感的には、「A_t[i, j]」は、エージェント「v_i」の役割変更がエージェント「v_j」のインセンティブをどの程度混乱させるかを測定します。大きな正の値は、強い相補性を示します (「v_i」が「v_j」の現在の役割をより価値のある役割に変更した場合)。大きな負の値は置換効果を示します (「v_i」が「v_j」と同じ役割に変化し、競合によって「v_j」の効用が減少する場合)。慣例により、対角線はゼロです: A_t[i, i] = 0。

The influence matrix captures three types of inter-agent effects: 1. Task complementarity. If agent v_i specializes in planning and agent v_j specializes in execution, v_i's planning output feeds v_j's execution pipeline. Changes in v_i's role disrupt this pipeline, creating a positive off-diagonal entry. 2. Resource competition. If v_i and v_j both specialize in the same role, they compete for the same task pool. Each additional agent in a role reduces the marginal return, creating negative entries. 3. Information effects. Agent v_i's role choice signals information about the state of the environment (e.g., if v_i switches to an auditing role, it may signal that error rates are high). This information affects v_j's utility estimates even if there is no direct task or resource interaction.

3.2 スペクトル半径と伝播ダイナミクス

Definition 5 (Spectral Radius). The spectral radius of A_t is: λ_max(A_t) = max{|λ| : λ ∈ spectrum(A_t)} where spectrum(A_t) is the set of eigenvalues of A_t. The spectral radius controls how influence perturbations propagate through the network. Consider a perturbation δρ_0 to the role assignment at time t = 0. After one round of best-response updates, the perturbation to agent utilities is approximately A_t · δρ_0. After k rounds, the cumulative perturbation is: δU_k ≈ A_t^k · δρ_0 The norm of this perturbation grows as ||δU_k|| ≤ ||A_t||^k · ||δρ_0||. For large k, this is dominated by the spectral radius: ||δU_k|| ∼ λ_max(A_t)^k · ||δρ_0|| If λ_max(A_t) < 1, perturbations decay exponentially: each round of best-response updates shrinks the displacement, and the system converges to a fixed point. If λ_max(A_t) > 1, perturbations grow exponentially: each round amplifies the displacement, and the system diverges. If λ_max(A_t) = 1, the system is marginally stable and requires higher-order analysis.

3.3 ガバナンスの減衰

Governance constraints act as a damping mechanism on influence propagation. When an agent faces an approval requirement before changing roles, the effective influence on that agent is reduced because the agent cannot immediately respond to the incentive change — it must wait for approval, during which time the incentive may have shifted again. When an agent must provide evidence for its role choice, the agent is less responsive to transient fluctuations in the utility landscape because evidence collection introduces a smoothing delay. Formally, governance with constraint density D_t reduces the effective influence matrix to: A_t^{eff} = (1 − D_t) · A_t The factor (1 − D_t) uniformly scales all influence strengths. This is a first-order approximation; in practice, different constraints affect different agents and different influence channels asymmetrically. The uniform scaling captures the aggregate effect: a fraction D_t of the action space is constrained, so a fraction D_t of influence-induced role changes are blocked. The effective spectral radius is: λ_max(A_t^{eff}) = (1 − D_t) · λ_max(A_t) The exact first-order contraction condition is therefore: (1 − D_t) · λ_max(A_t) < 1 which is equivalent to: λ_max(A_t) < 1 / (1 − D_t) This is the mathematically correct local stability condition for the linearized system. But enterprise architects usually need something stricter than bare contraction. A system with loop gain 0.99 is technically stable and still operationally brittle: convergence is slow, estimation error can flip the sign, and bursty or asynchronous updates can push the organization into oscillation. For that reason we define the Buffered Specialization Law: λ_max(A_t) < 1 − D_t Because 1 − D_t < 1 / (1 − D_t) for D_t ∈ (0, 1), this law is conservative. It guarantees contraction and preserves adaptation headroom, creating a region where specialization is not merely stable in theory but robust in operation.

4. 役割特化のダイナミクス

4.1 役割選択ルール

Each agent selects its role at each time step according to a utility-maximizing rule: Definition 6 (Role Selection). Agent v_i at time t selects: r_i(t+1) = argmax_r U_i(r | C_task, B_comm, D_t) where U_i is the role utility function, C_task is the current task context (available tasks, deadlines, priorities), B_comm is the communication bandwidth available to v_i, and D_t is the governance constraint density. This selection rule is myopic: the agent maximizes immediate utility without planning ahead. This is a deliberate simplification. In practice, agents may use multi-step lookahead (reinforcement learning policies), but the stability analysis applies to the best-response dynamics of the myopic rule, which forms the foundation for more sophisticated policies. If the myopic dynamics are unstable, no amount of lookahead can stabilize them without changing the influence structure itself.

4.2 ユーティリティ関数

The role utility function decomposes into three components: U_i(r) = α · Eff_i(r) + β · Impact_i(r) − γ · Cost_i(r, D_t) where α, β, γ > 0 are weighting parameters with α + β + γ = 1. Efficiency `Eff_i(r)`. This measures how well agent v_i's capabilities match role r's requirements. It is a function of the agent's knowledge capital K_t(v_i) and the skill profile of role r: Eff_i(r) = 1 − ||K_t(v_i) − K_req(r)||_2 / ||K_req(r)||_2 where K_req(r) is the required skill vector for role r. An agent whose knowledge profile exactly matches the role requirements achieves Eff = 1. An agent with no relevant skills achieves Eff ≤ 0. Impact `Impact_i(r)`. This measures the marginal contribution of agent v_i in role r to the overall organizational output. It depends on the current role distribution: if many agents are already in role r, the marginal impact of one more is low (diminishing returns). If no agents are in role r and there is demand for it, the marginal impact is high: Impact_i(r) = Demand(r) / (1 + n_r) where Demand(r) is the current task demand for role r and n_r is the number of agents currently in role r. This creates a natural load-balancing effect: excess agents in one role are drawn toward under-served roles. Cost `Cost_i(r, D_t)`. This measures the governance burden of agent v_i operating in role r under constraint density D_t. Higher constraint density increases the cost of every role because more actions require approval, evidence, or waiting: Cost_i(r, D_t) = D_t · (1 + SwitchCost(r, ρ_t(v_i))) where SwitchCost(r, ρ_t(v_i)) is the cost of switching from the current role ρ_t(v_i) to role r. This is zero if the agent is already in role r and positive otherwise, reflecting the governance overhead of role transitions (approval delays, re-certification, evidence requirements). The cost term creates a stickiness effect: under high D_t, agents are reluctant to change roles because the governance cost of switching is high. This is the direct mechanism through which governance damping operates at the agent level.

4.3 最良応答ダイナミクス

役割選択ルールを同時に適用する「N」個のエージェントのシステムにより、最良の応答ダイナミクスが定義されます。各タイム ステップで、すべてのエージェントはすべての役割に対するユーティリティを計算し、(ガバナンスの制約に従って) 最適なものに切り替えます。問題は、この力学が収束するかどうかです。 ρ_t ∈ ℛ^N を時間 t における役割割り当てベクトルとする。最良の応答マップは次のとおりです。 BR(ρ_t) = (argmax_r U_1(r | ρ_t, D_t), ..., argmax_r U_N(r | ρ_t, D_t)) 安定した役割割り当ては、「BR」の固定点、つまり「BR(ρ) = ρ」となるような役割ベクトル「ρ*」です。固定点では、他のすべてのエージェントの役割を考慮すると、どのエージェントもその役割を変更するインセンティブを持ちません。これはまさに役割選択ゲームのナッシュ均衡です。 固定点の存在は、混合戦略 (役割にわたる確率分布) に拡張すると、Brouwer の定理によって保証されます。問題は最善の対応かどうかだdynamic converges to this fixed point from arbitrary initial conditions. This is where the spectral radius condition enters.

4.4 平衡に近い線形化されたダイナミクス

固定点 ρ* の近くでは、最良応答ダイナミクスを線形化できます。 δρ_t = ρ_t − ρ* を平衡からの偏差とする。線形化された更新は次のとおりです。 δρ_{t+1} = J(BR) · δρ_t ここで、J(BR) は、ρ* における最良応答マップのヤコビアンです。ヤコビアンは、役割の割り当てにおける小さな変更が、すべてのエージェントの最良の応答の更新を通じてどのように伝播するかを捉えます。行列 J(BR) は、効用関数を通じて影響行列 A_t に直接関連付けられます。 J(BR)[i, j] = ∂(argmax_r U_i(r | ρ))[j] / ∂ρ_j 滑らかな効用近似 (不連続性を避けるために argmax がソフトマックスに置き換えられる) では、ヤコビアンは有効影響行列に比例します。 J(BR) ≈ (1 − D_t) · A_t 線形化された動的関数は、「J(BR)」のすべての固有値の法が 1 未満である場合にのみ収束します。それには、「(1 − D_t) · λ_max(A_t) < 1」が必要です。バッファされた特殊化condition λ_max(A_t) < 1 − D_t identifies the subset of that stable region with explicit governance reserve against perturbations in D_t, estimation error in A_t, and asynchronous update jitter.

5. The Stability Law

5.1 主定理

Theorem 1 (Exact Local Stability Law). Let G_t = (V, E_t, S_t, Π_t, R_t, D_t) be an agentic company with N agents, M roles, influence matrix A_t, and governance constraint density D_t ∈ (0, 1). Under Assumptions A1–A4 (stated below), the linearized best-response dynamic converges locally to a stable role assignment ρ* if and only if: (1 − D_t) · λ_max(A_t) < 1 Moreover, when the condition holds, convergence is exponential with rate r = 1 − (1 − D_t) · λ_max(A_t), and the number of steps to reach an ε-neighborhood of ρ* is at most: T(ε) = ⌈log(||δρ_0|| / ε) / (-log((1 − D_t) · λ_max(A_t)))⌉

前提: - A1 (Smooth Utilities)。 効用関数 U_i(r | ρ, D) は、すべての i、r、および D に対する役割分布 ρ で 2 回連続微分可能です。 - A2 (境界影響)。 影響行列は、t とは独立したいくつかの有限境界 B に対して ||A_t||_F ≤ B を満たします。ここで、||.||_F はフロベニウス ノルムです。 - A3 (非縮退ガバナンス)。 制約密度は、すべての t について 0 < D_t < 1 を満たします。完全に管理されていないシステムや完全に凍結されたシステムはありません。 - A4 (同期更新)。 すべてのエージェントは、各タイム ステップで自分の役割の選択を同時に更新します。非同期のバリアントについてはセクション 7 で説明します。

5.2 Proof Sketch

必要性 (`(1 − D)λ_max ≥ 1` は非収縮を意味します)。 (1 − D_t) · λ_max(A_t) ≥ 1 と仮定します。このとき、 |(1 − D_t) · λ| となるような固有値 λ を持つ A_t の固有ベクトル ξ が存在します。 1以上。小さな ε > 0 に対する摂動 δρ_0 = εξ を考えます。線形化された更新により次のことが得られます。 δρ_1 = (1 − D_t) · A_t · εξ = (1 − D_t) · λ · εξ 摂動はこの固有方向には縮小しないため、線形化されたマップは縮小することはできません。係数が厳密に 1 より大きい場合、摂動は指数関数的に増大し、平衡は局所的に不安定になります。 十分性 (`(1 − D)λ_max < 1` は収束を意味します)。 (1 − D_t) · λ_max(A_t) < 1 と仮定します。この場合、ヤコビアンの有効スペクトル半径は次の条件を満たします。 ρ(J(BR)) = (1 − D_t) · λ_max(A_t) < 1 縮小写像定理により、最良応答マップ BR は、レート (1 − D_t) · λ_max(A_t) をもつ ρ* の近傍の縮小です。のBanach fixed-point theorem guarantees local existence, uniqueness, and exponential convergence of the iterates ρ_t = BR^t(ρ_0) to ρ*. The stronger buffered specialization condition λ_max(A_t) < 1 − D_t is not the exact necessity threshold; it is a conservative design corollary. It carves out the portion of the stable region where governance retains explicit reserve against finite-sample uncertainty and operational perturbations.

5.3 直感

実際には 2 つのスカラー量が重要です。 - ループ ゲイン g = (1 − D) · λ_max(A)。これが正確な陣痛診断です。 g < 1 の場合、摂動は減衰します。 「g ≥ 1」の場合、摂動は持続するか増大します。 - バッファマージン δ_buffer = 1 − D − λ_max(A)。これは保守的な営業準備金です。 δ_buffer > 0 の場合、組織は収縮しているだけでなく、バ​​ッファされた特殊化エンベロープ内に快適に収まります。 これらの数量により、よくある混乱が解決されます。ガバナンスを強化すると、影響が弱まるため、常にループ ゲインが減少します。しかし、ガバナンスを強化すると自由も減り、生産的な適応に利用できるバッファーも縮小します。企業は数学的には安定していても (g < 1)、経営的には脆弱である (δ_buffer ≈ 0) 場合もあります。したがって、バッファの法則は単なる数学的な好奇心ではなく、ガバナンス設計ルールです。

6. Phase Diagram

6.1 The Four Regimes

厳密法則とバッファ法則を組み合わせたものにより、「(λ_max, D)」パラメータ空間が実質的に異なる 4 つの領域に分割されます。 フェーズ I: 停滞 (D > D_stag、通常は D_stag ≈ 0.7) このフェーズでは、ガバナンスの制約密度が非常に高いため、エージェントはその役割を効果的に適応できません。コスト項 γ · Cost(r, D_t) が支配的であり、ロールの切り替えに法外なコストがかかります。このシステムは、縮小という意味では非常に安定しているかもしれませんが、組織的には健全ではありません。 停滞フェーズで観察可能な指標: - 役割変更頻度 ≈ 0 - 需要の変化によりタスク完了率が低下 - ガバナンス間接費比率 > 60% - イノベーション率 ≈ 0 フェーズ II: バッファリングされた特殊化 (λ_max < 1 − D、D < D_stag を使用) これは望ましい動作体制です。エージェントには適応する十分な自由があり、ループ ゲインは快適に 1 未満であり、組織は明示的な情報を保持します。governance reserve. Perturbations are damped quickly, and specialization patterns remain legible under moderate shocks. Observable metrics in the Buffered Specialization phase: - Role distribution converges within 80–200 steps - Task completion rate > 90% - Role entropy stabilizes at an intermediate value - Perturbation recovery time remains bounded and predictable Phase III: Fragile Specialization (1 − D ≤ λ_max < 1 / (1 − D), with D < D_stag) Here the organization is still locally contractive, but only barely. It converges under the nominal model, yet small topology changes, delayed approvals, or asynchronous bursts can trigger long oscillation tails. This is the regime that often looks healthy in a dashboard until a mild stress event exposes the missing buffer. Observable metrics in the Fragile Specialization phase: - Convergence occurs but only after long settling times - Recovery time is highly variable across trials - Role entropy overshoots before settling - Alert frequency rises even though the system eventually reconverges Phase IV: Cascade ((1 − D)λ_max ≥ 1) In this phase, influence propagation dominates governance damping. When one agent changes roles, the utility perturbation propagates through the network, causing multiple other agents to change roles, which generates further perturbations, creating a cascade that does not contract. Observable metrics in the Cascade phase: - Role assignment entropy approaches its maximum - Task completion rate < 40% - Influence cascade length is effectively unbounded - Coordination capital C_t collapses toward zero

6.2 位相境界

Two boundaries matter: - The buffered operating boundary is the straight line λ_max = 1 − D. Points below this line lie in Buffered Specialization. - The exact contraction boundary is the hyperbola λ_max = 1 / (1 − D). Crossing this boundary moves the system from Fragile Specialization into Cascade. The boundary between Buffered Specialization and Stagnation is less sharp. We define the stagnation threshold as the constraint density D_stag beyond which the expected role change rate drops below one change per 1,000 agent-steps. Empirically, D_stag ≈ 0.70 for the utility function parameters used in our experiments, but this threshold depends on the specific values of α, β, γ and the switch-cost function. | Phase | Region | Behavior | Convergence | |-------|--------|----------|-------------| | Stagnation | D > D_stag | Frozen roles | Trivial | | Buffered Specialization | λ_max < 1 − D, D < D_stag | Robust convergent specialization | Exponential with healthy reserve | | Fragile Specialization | 1 − D ≤ λ_max < 1 / (1 − D), D < D_stag | Convergent but brittle | Slow / variance-sensitive | | Cascade | (1 − D)λ_max ≥ 1 | Oscillatory or divergent | None | The phase diagram reveals a fundamental trade-off. Increasing governance lowers loop gain and therefore helps contraction, but it also shrinks the operating buffer and can eventually push the organization into Stagnation. The practical optimum lies in the buffered region with D in the range [0.30, 0.55] — enough structure to prevent cascades, enough freedom to preserve adaptive specialization.

6.3 重大な行為

2 種類の重大な速度低下の問題: - 正確な縮小境界 付近では、ループ ゲインは次のように単一および収束時間スケールに近づきます。 T(ε) 〜 1 / (1 - (1 − D)λ_max) を ((1 − D)λ_max) → 1^- とすると - バッファ境界 付近では、バッファ マージンが消失し、運用上の差異が次のように増加します。 Var(Δρ) 〜 1 / (1 − D − λ_max)^2 最初の量は、収縮の数学的損失を捉えます。 2 つ目は、ガバナンス準備金の実質的な損失を捉えています。構造化されたトポロジ (MARIA 座標階層など) を持つネットワークでは、指数は異なる場合がありますが、正確な安定性とバッファリングされた動作の区別は引き続き重要です。

7. 収束解析

7.1 Main Convergence Result

定理 2 (指数関数的収束)。 正確な安定条件 (1 − D_t) · λ_max(A_t) < 1 の下では、期待される状態偏差は指数関数的に収束します。 lim_{t → ∞} E[||S_{t+1} − S_t||] = 0 さらに、収束率は次のようになります。 E[||S_{t+1} − S_t||] ≤ c_0 · ((1 − D) · λ_max)^t ここで、c_0 = ||S_1 − S_0|| は初期状態の変化の大きさです。 証明 状態ベクトル S_t(v_i) は、セクション 2.2 で定義された遷移関数 f に従って進化します。固定点 S* の近くでは、線形化されたダイナミクスはヤコビアンによって支配されます。 δS_{t+1} = J_f · δS_t ここで、「J_f」は状態ベクトルに関する状態遷移関数のヤコビアンです。重要な観察は、「J_f」が次のように分解されるということです。 J_f = J_self + (1 − D_t) · J_influence ここで、「J_self」は各エージェントの固有の状態ダイナミクス (安定したエージェントの場合は 1 未満に制限されます) をキャプチャし、「J_influence」はinter-agent influence terms mediated by A_t. When the combined loop gain remains below 1, the linearized operator is a contraction and the deviation norm decays geometrically. Exponential convergence follows from the submultiplicativity of the operator norm under iteration.

7.2 収束の必要条件

The convergence result requires four necessary conditions: 1. Individual Agent Stability. Each agent, considered in isolation (with fixed neighbor states), must converge to a stable state. This excludes agents with internally chaotic dynamics (e.g., oscillating reward estimates due to non-stationary exploration policies). Formally: ρ(J_self(v_i)) < 1 for all v_i. 2. Bounded Influence. The influence matrix must have bounded spectral radius. This is guaranteed by Assumption A2 (bounded Frobenius norm) and is satisfied in practice for any finite network with finite utility gradients. 3. Non-Zero Governance. The constraint density must be strictly positive: D_t > 0. A completely ungoverned system (D = 0) has stability condition λ_max < 1, which is generically violated in densely connected networks. Some governance is always necessary for stability in multi-agent systems. 4. Consistent Constraints. Governance constraints must be consistent: they must not create cycles where an agent is forced to change roles to satisfy one constraint and then forced to change back to satisfy another. Consistent constraints can be verified by checking that the constraint graph is acyclic.

7.3 Asynchronous Updates

定理 1 は同期更新を前提としています (仮定 A4)。実際には、エージェントは非同期に更新する場合があります。各エージェントは独自の速度で更新し、グローバル クロックではなくイベントによってトリガーされる可能性があります。 命題 1 (非同期安定性)。 正確な安定性条件 (1 − D_t) · λ_max(A_t) < 1 が成立する場合、最良応答動的は、すべてのエージェントが無限に頻繁に更新する非同期更新スケジュールの下で収束します。 プルーフ スケッチ。 非同期更新では、各ステップでの有効な更新演算子はエージェントのサブセットにのみ影響します。 'A_t' の主部分行列のスペクトル半径は最大でも 'λ_max(A_t)' (固有値インターレース) です。したがって、グローバル ループ ゲインが 1 未満の場合、各部分更新は縮小的なままであり、縮小化の構成自体が縮小化されます。収束は、非同期反復の標準理論に基づきます。paracontracting operators. The practical implication is that the exact gain condition governs mathematical convergence regardless of update ordering, while the buffered envelope remains the recommended target for robust production operation.

7.4 Convergence Speed

収束速度は、実効ループ ゲイン λ_{eff} = (1 − D) · λ_max によって制御されます。より高速なコンバージェンスには、次のいずれかが必要です。 1. 低い「λ_max」 — エージェント間の影響力が弱くなる (ネットワークが希薄になり、相補性が弱くなる) 2. 「D」が高い — より強力なガバナンスの減衰 (制約が増え、役割の変更が遅くなる) オプション 1 は設計上の選択です (ネットワーク トポロジ、ユーティリティ設計)。オプション 2 はガバナンス上の選択です。トレードオフは、より高い「D」によるより速い収縮は、適応性の低下と、ある点を超えると停滞を犠牲にして生じることです。 Planet-100 の実験 (セクション 10) では、バッファーされた特殊化分布の `ε = 0.01' 以内に収束するには、バッファーされた領域で 80 ～ 200 ステップが必要でした。脆弱な体制にあるシステムも収束しましたが、分散ははるかに大きく、整定時間ははるかに長くなりました。

8. Civilization Extension: 多層の安定性

8.1 Governance Layers

Agentic Companyは単独で活動するわけではありません。それは、規制の枠組み、業界標準、市場状況の中に存在し、また、MARIA OS のコンテキストでは、独自のガバナンス制約を課す文明シミュレーションの中に潜在的に存在します。これらの外部層により、安定性解析で考慮する必要がある追加の拘束密度が追加されます。 2 層システムを考えてみましょう。内部ガバナンス密度「D_company」を持つAgentic Companyが、外部ガバナンス密度「D_civ」を持つ文明 (または規制環境) 内で運営されています。企業レベルの制約により、個々のエージェントのアクション (承認ゲート、証拠要件、役割変更のクールダウン) が制限されます。文明レベルの制約は、企業全体の行動 (規制遵守、市場ルール、条約の義務) を制限します。

8.2 Effective Constraint Density

Definition 7 (Effective Constraint Density). For a two-layer governance system with company-level density D_company and civilization-level density D_civ, the effective constraint density is: D_eff = 1 − (1 − D_company)(1 − D_civ) Expanding: D_eff = D_company + D_civ − D_company · D_civ This formula has an intuitive interpretation. The freedom of an agent is the fraction of its action space that is unconstrained. Company-level governance leaves a fraction (1 − D_company) unconstrained. Civilization-level governance further constrains a fraction D_civ of the remaining freedom. The residual freedom is (1 − D_company)(1 − D_civ), so the effective density is 1 − (1 − D_company)(1 − D_civ). Note that D_eff > max(D_company, D_civ) whenever both are non-zero: multiple governance layers always increase effective constraint density. This has important implications for multi-layer stability.

8.3 多層安定性の法則

Theorem 3 (Multi-Layer Exact Stability). For a two-layer governance system, the exact contraction condition becomes: (1 − D_company)(1 − D_civ) · λ_max(A_t) < 1 Equivalently: λ_max(A_t) < 1 / (1 − D_eff) Adding a civilization layer with D_civ = 0.2 to a company with D_company = 0.3 gives D_eff = 0.44, so the exact contraction condition becomes λ_max < 1 / 0.56 ≈ 1.79. Corollary 1 (Buffered Multi-Layer Operation). The conservative buffered envelope becomes: λ_max(A_t) < 1 − D_eff = (1 − D_company)(1 − D_civ) This is the recommended production condition for robust specialization across layers. Corollary 2. Additional governance layers reduce the buffered operating region even while they improve pure damping. In other words, layered regulation can make contraction easier but still push the organization toward stagnation by exhausting adaptation headroom. Corollary 3. There exists a maximum number of governance layers L_max beyond which buffered specialization is no longer possible for a given λ_max: L_max = ⌈log(λ_max) / log(1 − D_avg)⌉ where D_avg is the average constraint density across layers. For typical values (λ_max = 0.7, D_avg = 0.2), this gives L_max ≈ 3 layers before the system exits the buffered region even though exact contraction may still hold.

8.4 文明シミュレーションマッピング

In the MARIA OS Civilization simulation (Section 10), nations correspond to agentic companies, and the civilization-level governance includes: - Market rules (free market land economy, trading constraints) - Election cycles (political transitions that reset governance parameters) - Treaty obligations (inter-nation constraints on resource allocation) - LOGOS AI advisor recommendations (soft constraints that influence but do not mandate policy) Each of these contributes to D_civ. The effective constraint density for a nation is: D_eff(nation) = 1 − (1 − D_internal)(1 − D_market)(1 − D_political)(1 − D_treaty)(1 − D_LOGOS) where each sub-density captures a different governance channel. This decomposition allows the Civilization simulation to model how different governance structures (authoritarian vs. democratic, regulated vs. free-market) produce different 実効拘束密度、したがって異なる安定性特性。

9. MARIA OS Implementation

9.1 アーキテクチャのマッピング

The stability framework maps directly onto the MARIA OS architecture. The MARIA coordinate system G.U.P.Z.A (Galaxy, Universe, Planet, Zone, Agent) provides the graph structure G_t = (V, E_t). The influence edges E_t correspond to the communication and dependency links between agents within and across zones. The governance constraint density D_t is computed from the active gate configurations, approval policies, and evidence requirements at each hierarchical level. The mapping is hierarchical. Each level in the coordinate system contributes to the effective constraint density: ``yaml # MARIA OS Constraint Density Configuration galaxy: D_tenant: 0.10 # Enterprise-wide policies constraints: - global_compliance_standards - tenant_budget_limits - cross_universe_approval_gates universe: D_business_unit: 0.12 # Business unit policies constraints: - unit_spending_authority - role_assignment_policies - inter_planet_coordination_rules planet: D_domain: 0.15 # Domain-specific governance constraints: - domain_expertise_requirements - quality_gate_thresholds - audit_frequency_policies zone: D_ops: 0.08 # Operational constraints constraints: - task_priority_rules - resource_allocation_caps - agent_cooldown_periods agent: D_individual: 0.05 # Per-agent constraints constraints: - role_change_cooldown - evidence_requirements - escalation_triggers # Effective constraint density for agent at G1.U2.P3.Z1.A4: # D_eff = 1 - (1-0.10)(1-0.12)(1-0.15)(1-0.08)(1-0.05) = 0.41 # Exact loop gain check: g = D_free * lambda_max(A_t)

9.2 Gate Engine as Governance Density Controller

MARIA OS ゲート エンジン (lib/engine/decion-pipeline.ts および lib/engine/responsibility-gates.ts に実装) は、ガバナンス制約密度を制御するための主要なメカニズムです。各ゲート タイプは D_t に寄与します。 - 承認ゲートは、人間の承認が必要なアクションを制限することで「D」を増加させます。アクティブな承認ゲートの数を合計アクション数で割った値は、「D_t」に直接寄与します。 - 証拠ゲートは、エージェントが行動する前に証拠を収集して提示することを要求することで、「D」を増加させます。それぞれの証拠要件により、アクション スペースに制約が追加されます。 - 予算ゲートは、役割の変更やタスクの実行に利用できる財源を制限することで「D」を増加させます。 - クールダウン ゲートは、エージェントが役割を変更できる頻度に一時的な制約を課すことにより、「D」を増加させます。 ゲート エンジンは、すべての組織レベルにわたる「D_t」のリアルタイム監視を提供します。enabling administrators to tune governance intensity to maintain the stability condition. When λ_max is estimated to be rising (e.g., due to increased network connectivity or stronger complementarities), D_t can be increased by activating additional gates. When D_t is too high (risk of Stagnation), gates can be relaxed. This is the core operational insight: governance gates are not bureaucratic checkpoints — they are control inputs in a dynamical system. Tuning gate parameters is equivalent to choosing the damping coefficient in a control loop.

9.3 Evidence Layer as Influence Damper

The MARIA OS evidence layer (implemented in lib/engine/evidence.ts) serves a dual purpose in the stability framework. First, it provides the audit trail required for accountability. Second, it acts as an influence damper by introducing delay between an agent's decision to change roles and the execution of that change. When an agent decides to switch roles, the evidence requirement forces it to: 1. Collect evidence supporting the switch (task demand data, capability match scores, coordination impact estimates) 2. Bundle the evidence into an evidence package 3. Submit the package for review (automated or human) 4. Wait for approval before executing the switch This delay prevents agents from responding instantaneously to utility perturbations. The smoothing effect is analogous to a low-pass filter: high-frequency role oscillations are filtered out because the evidence collection このプロセスには発振周期よりも時間がかかります。持続的で真の効用の変化のみが証拠フィルターを通過し、実際の役割の変化を生み出します。

9.4 Anomaly Detection: Stability Monitoring

MARIA OS は、分析エンジン (lib/engine/analytics.ts) を通じてリアルタイムの安定性モニタリングを実装します。追跡される主要な指標は次のとおりです。 1. スペクトル半径の推定。 影響行列 'A_t' は、観察されたエージェントの相互作用 (役割の変化、ユーティリティの摂動) から推定されます。スペクトル半径は、経験的影響行列に対するべき乗反復を使用して一定の間隔で計算されます。 2. 制約密度測定 「D_t」は、アクティブなゲート構成と現在のアクション空間サイズから計算されます。 3. ループ ゲイン。 正確な安定性診断 g_t = (1 − D_t) · λ_max(A_t) がリアルタイムで追跡されます。 g_t が 1 に近づくと、システムは収縮の喪失に近づいています。 4. バッファ マージン。 保守的な動作予備量 δ_buffer = 1 − D_t − λ_max(A_t) が並行して追跡されます。 「δ_buffer」がゼロに近づくと、組織は脆弱な体制に入ります。if contraction still holds. 5. Role Change Frequency. A sudden increase in role change frequency is an early indicator of approaching the cascade boundary. 6. Perturbation Recovery Time. The time for the system to return to steady state after a perturbation is measured. Increasing recovery time indicates critical slowing down near the boundary. When either g_t > 0.95 or δ_buffer < 0.10, the anomaly detection system triggers an alert and can activate additional governance gates to increase D_t. This creates a self-stabilizing feedback loop: instability triggers governance, governance restores contraction and, where possible, rebuilds operating buffer.

9.5 ユニバースダッシュボードの統合

安定性メトリックは、ダッシュボード データ プロバイダー (lib/contexts/dashboard-data-context.tsx) を通じて MARIA OS Universe ダッシュボードに表示されます。各ユニバース ビューには次のものが含まれます。 - ループ ゲイン ゲージ: 「g_t = (1 − D_t)λ_max(A_t)」のリアルタイム表示。緑 (安全)、オレンジ (警告)、赤 (クリティカル) に色分けされます。 - バッファー ゲージ: δ_buffer = 1 − D_t − λ_max(A_t) の表示。組織がバッファーされているか脆弱であるかを示します。 - フェーズインジケーター: (λ_max, D) 座標に基づく現在の動作フェーズ (停滞 / バッファリングされた特殊化 / 脆弱な特殊化 / カスケード)。 - 収束タイマー: 定常状態の役割配分までの推定ステップ。T(ε) = log(||δρ_0|| / ε) / (-log(λ_{eff})) から計算されます。 - 役割分布グラフ: エージェントから役割への割り当てのリアルタイム ヒストグラム。平衡分布への収束を示します。 - 影響ヒートマップ:Visualization of the N × N influence matrix A_t, highlighting the dominant eigenvector (the direction of maximum instability).

10. プラネット-100の実験

10.1 実験のセットアップ

We validate the stability law using the MARIA OS Planet-100 simulation environment (app/experimental/planet-100/). Planet-100 simulates a self-organizing society of autonomous agents, providing a controlled testbed for studying emergent role specialization. Configuration: | Parameter | Value | |-----------|-------| | Agent Count (N) | 111 | | Role Types (M) | 10 (Planner, Executor, Auditor, Negotiator, Researcher, Communicator, Analyst, Guardian, Optimizer, Coordinator) | | Simulation Steps | 1,000 per trial | | Trials per Configuration | 50 | | Utility Weights | α = 0.4, β = 0.35, γ = 0.25 | | Influence Topology | Scale-free (preferential attachment, m = 3) | | PRNG | Deterministic seeded (mulberry32) for reproducibility | The 111 agents correspond to the Planet-100 population (100 primary agents plus 11 governance observers). The 10 role types represent the functional roles identified in the MARIA OS agent taxonomy. The scale-free topology produces a power-law degree distribution, which is representative of real organizational networks where some agents (hubs) have disproportionate influence.

10.2 パラメータスイープ

We conduct two parameter sweeps: Sweep 1: Constraint Density `D` ∈ [0.05, 0.90] at fixed topology. The influence matrix A_t is fixed (same topology for all trials), and D is varied in increments of 0.05. For each value of D, we run 50 trials with random initial role assignments and measure: - Whether the role distribution converges (defined as ||S_{t+1} − S_t|| < 0.01 for 20 consecutive steps) - The convergence time (number of steps to reach the convergence criterion) - The final role entropy H(ρ̂) = −Σ_k ρ̂(k) · log(ρ̂(k)) - The task completion rate at steady state Sweep 2: Spectral Radius `λ_max` ∈ [0.2, 1.5] at fixed `D`. The constraint density is fixed at D = 0.40, and the influence matrix is scaled to achieve different spectral radii. Specifically, we compute the natural influence matrix A_0 from the scale-free topology and scale it: A_t = s · A_0 where s = λ_target / λ_max(A_0). This preserves the structure of the influence network while controlling its strength. For each value of λ_max, we run 50 trials and measure the same convergence metrics.

10.3 バッファリングされた境界の再生

The combined results from both sweeps produce an empirical phase diagram in the (λ_max, D) plane. For the purposes of deployment guidance, we classify each point as buffered or non-buffered using the conservative operating boundary λ_max = 1 − D, while separately tracking whether the exact contraction condition is still satisfied. The empirical buffered boundary closely matches the theoretical prediction. Specifically: | Metric | Value | |--------|-------| | Buffered boundary classification accuracy | 96.8% | | Mean absolute error of buffered boundary | 0.032 | | False buffered (predicted buffered, actually fragile/cascade) | 1.4% | | False non-buffered (predicted fragile, actually buffered) | 1.8% | The 3.2% misclassification rate occurs at points very close to the boundary (|λ_max − (1 − D)| < 0.05), where finite-time effects and the discreteness of the agent population cause deviations from the infinite-population, continuous-time theory. The false buffered rate (1.4%) represents configurations that appear robust in early steps but are actually fragile under longer horizons. The false non-buffered rate (1.8%) represents configurations where nonlinear saturation provides a little more damping than the linear buffered model predicts.

10.4 収束速度の結果

Within the buffered phase, the convergence speed varies systematically with the buffer margin δ_buffer = 1 − D − λ_max: | Buffer Margin δ_buffer | Mean Convergence Steps | Standard Deviation | |---------------------------|------------------------|--------------------| | 0.05 (near buffered boundary) | 198 | 47 | | 0.10 | 142 | 31 | | 0.20 | 103 | 22 | | 0.30 | 87 | 18 | | 0.40 | 78 | 14 | | 0.50 | 65 | 11 | The convergence time scales approximately as T ∼ 1/δ_buffer inside the buffered regime, while the variance drops rapidly as the buffer grows. This is precisely why the buffered law is operationally useful: it predicts not just whether the system settles, but how reliably it settles.

10.5 最適なガバナンス密度

We define the optimal governance density D* as the value that maximizes a combined objective: Objective(D) = TaskCompletion(D) · Adaptivity(D) · Buffer(D) where: - TaskCompletion(D) is the steady-state task completion rate (decreasing in D due to governance overhead) - Adaptivity(D) is the speed of adaptation to demand shocks (decreasing in D due to role-change friction) - Buffer(D) is max(0, 1 − D − λ_max) The product peaks at D* ∈ [0.30, 0.55], with the exact optimum depending on the weighting of the three factors. For equal weighting, D* ≈ 0.40. This aligns with the intuition that moderate governance produces the best outcomes: enough structure to prevent cascades, enough freedom to enable adaptation. At the optimal density, the system exhibits: - Task completion rate: 93.4% (vs. 97.1% at D = 0.10 and 78.2% at D = 0.70) - Perturbation recovery time: 23 steps (vs. 8 steps at D = 0.10 and 450 steps at D = 0.70) - Role entropy: 2.1 (out of log(10) = 2.3, indicating diversified but not uniform specialization) - Buffer margin: 0.22 (comfortable reserve from the fragile regime)

11. ディスカッション

11.1 制御入力としてのガバナンス

この文書の中心的なメッセージは、ガバナンスはオーバーヘッドではなく、制御入力であるということです。制御理論のアナロジーでは、影響行列 A_t はプラントを表し、ガバナンス密度 D_t は減衰入力を表し、ループ ゲイン g_t = (1 − D_t)λ_max(A_t) は正確な収縮テスト、バッファ マージン δ_buffer = 1 − D_t − λ_max(A_t) は動作予備量です。制御エンジニアが裸の安定性と健全なゲインマージンを区別するのと同じように、エンタープライズアーキテクトは正確な収束とバッファされた特殊化を区別する必要があります。 この視点は、エージェント会社の設計で繰り返される議論、つまりガバナンスを最小化する必要があるのか​​ (エージェントの自律性と速度を最大化するため)、それとも最大化すべきなのか (リスクを最小化するため) を解決します。答えはどちらでもない。ループ ゲインを 1 未満に保ち、理想的には正の動作バッファーを維持するようにガバナンスを調整する必要があります。ガバナンスが少なすぎるallows cascading instability. Too much governance suppresses the adaptive capacity that justifies using agents in the first place.

11.2 Implications for Enterprise Architecture

The stability framework has several practical implications for enterprise architects designing agentic companies: Network Topology. The spectral radius λ_max is a property of the influence network. Densely connected networks with strong complementarities have high λ_max and therefore erode the operating buffer quickly. Architects can reduce λ_max by modularizing the organization into loosely coupled zones, standardizing interfaces between zones, and using the MARIA coordinate hierarchy to localize propagation. Governance Tuning. The constraint density D should be monitored and adjusted in real time, not set once at design time. Environmental changes alter the influence matrix, changing both loop gain and operating buffer. The MARIA OS gate engine provides the mechanism for this dynamic tuning. Scaling Laws. As the agent population N grows, the spectral radius of ランダム影響行列は、疎ネットワークの場合は「O(√N)」、密ネットワークの場合は「O(N)」としてスケールされます。したがって、大規模なAgentic Companyには、ガバナンスを強化するか、影響力を実質的に希薄にしたトポロジが必要です。 MARIA 座標階層は、ほとんどの影響をゾーン内に局所化し、ゾーン間のチャネルを慎重に管理することにより、自然なスケーリング メカニズムを提供します。

11.3 Relationship to Existing Frameworks

The framework connects to several established traditions: Game Theory. The stable role assignment ρ* is a Nash equilibrium of the role-selection game. The exact condition (1 − D)λ_max < 1 characterizes local convergence, while the buffered envelope λ_max < 1 − D gives a stronger, governance-oriented sufficient condition. Control Theory. The exact loop-gain test is the natural discrete-time contraction criterion for linearized feedback systems. The buffered envelope plays the role of a design margin rather than a bare feasibility threshold. Statistical Physics. The buffered boundary behaves like an order-disorder frontier, while the exact contraction boundary captures the onset of true amplification. The distinction mirrors the difference between nominal equilibrium and robust phase stability. Organizational Theory. The framework formalizes Ashby's Law of Requisite Variety: governance must supply enough constraint structure to absorb the variety generated by inter-agent influence, without over-constraining the organization into stagnation.

11.4 制限事項

このフレームワークには、将来の作業の方向性を示唆するいくつかの制限があります。 1. 線形近似。 正確な安定性の法則は、平衡に近い最良応答ダイナミクスの線形化から導出されます。収縮境界付近では、非線形効果が顕著になります。 2. 定常仮定 影響行列 'A_t' は、収束時間に対してゆっくりと変化すると仮定します。急速に変化する環境では、ループ ゲインとバッファの両方を継続的に追跡する必要があります。 3. 均一なガバナンス。 このモデルは、すべてのエージェントにわたって均一なガバナンス密度「D」を想定しています。実際には、エージェントが異なれば、ガバナンスの強度も異なります。フレームワークを異種密度に拡張するには、スカラー 'D' ではなく行列値の減衰係数が必要です。 4. 役割空間の離散性 連続分析 (スムーズ ユーティリティ、ヤコビアン) は、discrete role-selection problem. For small role spaces (M < 5), discrete effects may dominate. 5. No learning dynamics. The framework assumes fixed utility functions. Incorporating online learning, exploration, and adaptive topology formation remains a major open problem.

11.5 今後の方向性

Several extensions of the stability framework merit investigation: Adaptive Governance. Rather than setting D manually or through threshold-based rules, use online control to optimize both loop gain and buffer in real time. Heterogeneous Stability. Extend the framework to account for heterogeneous agents with different influence strengths, governance burdens, and utility functions. The generalized condition would involve the spectral radius of a damping-weighted influence matrix diag(1 − D_i) · A. Temporal Governance. Study how governance density should vary over the lifecycle of an agentic company. Early-stage exploration and mature operations may require different target buffers. Multi-Objective Stability. In practice, multiple stakeholders care about different kinds of stability (financial, operational, reputational, ethical). A multi-objective framework would 複数の影響マトリックスにわたる同時バッファ管理が必要です。

12. 結論

This paper has established a formal stability framework for agentic companies — enterprises in which autonomous AI agents self-organize into functional roles through decentralized utility maximization. The central result is a two-level criterion: - Exact contraction: (1 − D)λ_max(A) < 1 - Buffered specialization: λ_max(A) < 1 − D The first inequality determines whether the linearized organization actually contracts. The second identifies the portion of that stable region with usable governance reserve. Together they partition the design space into Stagnation, Buffered Specialization, Fragile Specialization, and Cascade. We have shown that this framework extends to multi-layer systems through the effective constraint density formula D_eff = 1 − ∏(1 − D_l). The MARIA OS implementation provides concrete tools for monitoring and controlling both loop gain and operating buffer: the gate engine controls D, the evidence layer damps high-frequency oscillations, the analytics engine tracks λ_max, g_t, and δ_buffer, and the dashboard surfaces these metrics to human operators. Experimental validation on the Planet-100 simulation with 111 agents and 10 role types confirms the buffered operating boundary with 96.8% classification accuracy. The optimal governance density range D ∈ [0.30, 0.55] produces buffered specialization within 80–200 convergence steps, with enough reserve to absorb perturbations without slipping into fragility. The deeper insight is that governance and autonomy are not opposing forces. Governance supplies the damping that makes autonomy sustainable, but only up to the point where adaptation headroom disappears. For enterprise architects designing AI-native organizations, the design prescription is now precise: build the influence graph, estimate λ_max, keep loop gain below 1, and target a positive operating buffer. Stability is not the absence of change. It is the presence of convergence with reserve.

参考文献

1. Ashby, W. R. (1956). An Introduction to Cybernetics. Chapman & Hall. 2. Barabási, A.-L. & Albert, R. (1999). Emergence of scaling in random networks. Science, 286(5439), 509–512. 3. Brooks, F. P. (1975). The Mythical Man-Month. Addison-Wesley. 4. Busoniu, L., Babuska, R., & De Schutter, B. (2008). A comprehensive survey of multiagent reinforcement learning. IEEE Transactions on Systems, Man, and Cybernetics, 38(2), 156–172. 5. Chandler, A. D. (1962). Strategy and Structure. MIT Press. 6. Daskalakis, C., Goldberg, P. W., & Papadimitriou, C. H. (2009). The complexity of computing a Nash equilibrium. SIAM Journal on Computing, 39(1), 195–259. 7. Fudenberg, D. & Tirole, J. (1991). Game Theory. MIT Press. 8. Horn, R. A. & Johnson, C. R. (2012). Matrix Analysis. Cambridge University Press. 9. Lowe, R., Wu, Y., Tamar, A., Harb, J., Abbeel, P., & Mordatch, I. (2017). Multi-agent actor-critic for mixed cooperative-competitive environments. NeurIPS. 10. Mintzberg, H. (1979). The Structuring of Organizations. Prentice-Hall. 11. Monderer, D. & Shapley, L. S. (1996). Potential games. Games and Economic Behavior, 14(1), 124–143. 12. Nash, J. F. (1950). Equilibrium points in n-person games. Proceedings of the National Academy of Sciences, 36(1), 48–49. 13. Newman, M. E. J. (2010). Networks: An Introduction. Oxford University Press. 14. Nyquist, H. (1932). Regeneration theory. Bell System Technical Journal, 11(1), 126–147. 15. Olfati-Saber, R., Fax, J. A., & Murray, R. M. (2007). Consensus and cooperation in networked multi-agent systems. Proceedings of the IEEE, 95(1), 215–233. 16. Ozdaglar, A., & Menache, I. (2011). Network Games: Theory, Models, and Dynamics. Morgan & Claypool. 17. Simon, H. A. (1955). A behavioral model of rational choice. Quarterly Journal of Economics, 69(1), 99–118. 18. Strogatz, S. H. (2015). Nonlinear Dynamics and Chaos. Westview Press. 19. Sutton, R. S. & Barto, A. G. (2018). Reinforcement Learning: An Introduction. MIT Press. 20. Taylor, F. W. (1911). The Principles of Scientific Management. Harper & Brothers. 21. Urwick, L. F. (1956). The manager's span of control. Harvard Business Review, 34(3), 39–47. 22. Vickrey, D. & Koller, D. (2002). Multi-agent algorithms for solving graphical games. AAAI. 23. Weiss, G. (Ed.). (2013). Multiagent Systems. MIT Press. 24. Wooldridge, M. (2009). An Introduction to MultiAgent Systems. John Wiley & Sons. 25. Young, H. P. (2004). Strategic Learning and Its Limits. Oxford University Press.