The hidden bottleneck in AI-native orgs: skills governance, not agents

The speaker argues that as organizations scale AI adoption, the critical missing piece is not better agents or larger context windows but a structured governance layer for 'skills' — reusable, modular, deterministic tasks that codify organizational expertise. Without this, teams generate duplication, inconsistent quality, and rising token costs.

Skills are positioned as the new unit of executable knowhow, analogous to microservices in the last decade. They require design principles: reusability, discoverability, portability across harnesses, composability, consistency, and cost efficiency. The core insight is that workflows rely on skills for determinism, not just on hooks, MCP servers, or subagents.

When skills are ungoverned, the simulation shows uneven productivity, high costs, and low quality across teams. By centralizing skills in a registry with versioning, metadata, access control, and evaluation, organizations can level-set performance. The speaker warns that auto-evolving skills without governance will amplify the problem.

The talk is largely prescriptive and draws parallels to past microservices and domain-driven design patterns. It offers a clear organizational playbook but lacks concrete benchmarks or real-world case studies demonstrating ROI, leaving some claims about productivity uplift as plausible but unproven.