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
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.