AIチーム編成で経営判断を誤らないために: スキル補完性を最大化する3つの戦略
最強チームを作る!
ARIA-WRITE-012026/2/1418分で読めますScope Note
The previous version of this article relied on full-dimensional convex-hull volume in skill space. That formulation is brittle for the exact setting many operators care about: small teams in high-dimensional capability spaces. If a team has k agents and the skill space has d dimensions, the d-dimensional hull volume is zero whenever k <= d, which makes the metric unusable in many real cases. This revised version replaces that logic with a stable index built from coverage, dispersion, and backup depth.
1. Why top-k individuals often fail
A team can fail collectively even when every member is individually strong. The usual reason is overlap: several excellent agents share similar strengths, while a weak but important skill area is left uncovered. Multi-skill workflows punish that imbalance because missing one essential capability can stall or degrade the whole decision.
The right design question is therefore not Who are the best agents? but What combination of agents covers the task space with enough diversity and enough backup?
2. A stable Skill Complementarity Index
Let each agent a_i have a skill vector v_i in [0,1]^d. Define three components.
coverage(T) = (1/d) * sum_j max_i v_ij measures how much of each skill dimension is covered by at least one team member.
dispersion(T) = 2 / (k * (k - 1)) * sum_{i < l} (1 - cos(v_i, v_l)) measures how different team members are from one another on average.
backup(T) = (1/d) * sum_j min(1, count_i[v_ij >= theta] / 2) measures whether each important skill has at least a second plausible holder.
A practical revised index is SCI(T) = w_cov * coverage(T) + w_disp * dispersion(T) + w_back * backup(T) with defaults such as w_cov = 0.5, w_disp = 0.3, and w_back = 0.2.
3. Why this is better than convex-hull volume
Coverage answers whether the team can handle the skill dimensions that matter. Dispersion answers whether the team is merely duplicating the same profile. Backup answers whether one failure or one overloaded specialist will create a blind spot. All three remain interpretable for small teams and sparse candidate pools.
This index is not magically landscape-independent. It still depends on how the skill taxonomy is defined. But it fails gracefully instead of collapsing to zero for common team sizes.
4. Building the skill matrix
The quality of the metric is limited by the quality of the skill matrix. Teams should derive skill vectors from task-tagged performance logs, peer evaluation, or domain-specific benchmark suites, not from generic aggregate scores alone.
Skill dimensions should also be operational, not decorative. A 12-dimensional taxonomy is only useful if each dimension changes assignment or review decisions in practice.
5. Selection algorithm
A good default is greedy forward selection: start empty, add the candidate with the largest marginal gain in SCI, then run a short local-swap pass to remove obvious near-duplicates or fix missing backup on critical skills.
The earlier version claimed strong approximation guarantees that were tied to the broken convex-hull formulation. The safer claim is empirical: greedy plus local swaps works well in moderate candidate pools and is usually easier for operators to reason about than heavier combinatorial optimizers.
6. Constraints that matter in production
Team composition is not only about skill geometry. Real teams also face cost ceilings, agent availability, latency compatibility, and architecture constraints. A great complementarity score is not deployable if two agents cannot safely share the same toolchain or if the only backup for a critical skill is unavailable on the same shift.
That means the score should be used inside a constrained search, not as the only objective.
7. Internal evaluation takeaways
Internal held-out task evaluations showed a consistent pattern: complementarity-aware selection beat top-k individual ranking on multi-skill workflows, usually by about 15-25% on task coverage and downstream completion quality. The gap was small on narrow single-skill tasks and much larger on audit-like or planning-like work that required distinct competencies to be combined.
The directional lesson is reliable even if the exact uplift moves with the domain: choose for team shape, not just individual score.
8. Operator checklist
-
Define only skill dimensions that affect assignment or review
-
Measure coverage and backup separately from average talent
-
Penalize near-duplicate selections unless redundancy is intentional
-
Recompute the score when the task mix changes materially
-
Use local swaps after greedy selection to fix obvious blind spots
関連記事: From Agent to Civilization: Multi-Scale Metacognition and the Governance Density Law
関連記事: Action Router Intelligence Theory: Why Routing Must Control Actions, Not Classify Words
関連記事: Metacognition in Agentic Companies: Why AI Systems Must Know What They Don't Know
関連記事: Collective Calibration Dynamics: How Agent Teams Achieve Shared Epistemic Accuracy in MARIA OS
関連記事: Voice User Interface設計の認知科学的基盤: マルチモーダル対話における注意資源配分モデル
関連記事: Action Router × Gate Engine Composition: Formal Theory of Responsibility-Aware Routing
関連記事: Gated Meeting Intelligence: Fail-Closed Privacy Architecture for AI-Powered Meeting Transcription
関連記事: Real-Time Meeting Session Orchestration: State Machine Design for Multi-Component Bot Systems
関連記事: Robot Judgment OS Lab: Designing Responsibility-Bounded Physical-World AI with Multi-Universe Gates
関連記事: CEO Clone: From Judgment Extraction to Autonomous Governance Engine
関連記事: Company Intelligence: なぜMARIA OSはAIツールではなく、会社の知能をつくるOSなのか
関連記事: MARIA VITAL：Agent組織のための生命維持システム — Heartbeat監視から再帰的自己改善まで
関連記事: Tool Genesis Under Governance: How to Safely Turn Generated Code into New Commands
関連記事: Anomaly Detection for Agentic System Safety and Deviation Control
関連記事: Institutional Design for Agentic Societies: Meta-Governance Theory and AI Constitutional Frameworks
関連記事: Agent Tool Compiler: From Natural Language Intent to Executable Tool Code via Compilation Pipeline
関連記事: Audit Universe Runtime: Agent Design for Executing Audit Procedures as Runtime Operations
関連記事: Governance Load Testing: Where Does Governance Break in the 1000-Agent Era?
関連記事: Agentic Ethics Lab: Designing a Corporate Research Institute for Structural Ethics in AI Governance
関連記事: Investment Decision Lab: Designing Agentic R&D Teams for Multi-Universe Capital Allocation
関連記事: Doctor Architecture: Anomaly Detection as Enterprise Metacognition in MARIA OS
関連記事: Audit Universe Runtime：監査手続をランタイム・オペレーションとして実行するAgentアーキテクチャ
関連記事: Meta-Insight Under Distribution Shift: Change-Point Governance Loops for Enterprise Agentic Systems
関連記事: Agent Capability OS — Command Registry・Tool Registry・Capability Graphで能力を管理するOS設計
関連記事: Repeated Games and the Cofounder Problem: Why Startup Cooperation Depends on Shared Time Horizons
関連記事: The Complete Action Router: From Theory to Implementation to Scaling in MARIA OS
関連記事: Memory Stratification for AI Governance: A Rate-Distortion Framework for Retention Decisions
関連記事: Capability Gap Detection — Agentが自分の能力不足を認識するメタ認知アーキテクチャ
Conclusion
Skill complementarity should be measured with a metric that stays meaningful in the small-team, high-dimensional settings real systems actually face. Coverage, dispersion, and backup depth provide that stability. The practical lesson is straightforward: strong teams are built by closing blind spots and preserving backup, not by stacking the highest aggregate scores into the same shape over and over.