Descriptions:
Imad Touil, Distinguished Engineer at QuantumBlack (McKinsey’s AI division), delivers a framework-level talk on structuring AI capabilities around reusable “skills” — modular, portable, composable units of organizational know-how that agents can discover and execute at runtime. The central argument is that most teams building with agents have mastered individual product increments but have not solved how to govern and scale skills across an entire organization.
Touil opens with an overview of what he calls the agentic software stack: an inner loop handling context management, tool and MCP integration, memory, and a skills loader, wrapped by an outer loop of workflows, sub-agents, and MCP servers. Supporting components include an MCP gateway, model gateway, and a knowledge graph that abstracts an organization’s codebase and IT systems. He frames the real challenge not as building a single feature but as orchestrating the full software delivery lifecycle — product strategy, market research, discovery, data preparation, and implementation — with agents operating coherently across all of it.
The core design principles for skills mirror microservices thinking: reusable, modular, discoverable, portable across harnesses (Claude Code, Cursor), specialized, composable, consistent, and cost-efficient. Touil specifically positions skills as a solution to context-window bloat through progressive disclosure — surfacing only the right skills at the right moment. A regulatory compliance example demonstrates composable skills for GDPR, data retention policies, and disclosure standards, producing deterministic audit reports at runtime.
📺 Source: AI Engineer · Published August 28, 2026
🏷️ Format: Deep Dive
