← back
AI-Native Organisations Run on Skills: How to Structure and Scale Them — Imad Touil, QuantumBlack
Takeaway
Scale agent adoption by managing skills as shared software assets that encode the organization's real workflows.
Summary
- An agentic stack combines an inner execution harness with an outer workflow layer of skills, subagents, MCP servers, and hooks.
- Enterprise delivery spans strategy, discovery, data preparation, implementation, platform operations, and optimization, requiring multiple specialized workflows.
- Skills carry organization-specific know-how while gateways, sandboxes, registries, and context infrastructure support execution at scale.
- The speaker applies familiar software design principles to skills, emphasizing reusable, modular, discoverable components and organization-wide maintenance.
agent-skillsenterprise-aiworkflow-design
Original description
Imad Touil opens with a show of hands. Who has built a skill? Most of the room. Who shares them with their team? Fewer. Who governs and maintains them across the organization? A handful. That collapsing sequence is the talk, because he argues skills are where an organization's actual know how ends up living. Of the four things inside a workflow, hooks only fire on events, sub agents mostly exist to protect the context window, and almost nobody writes their own MCP servers. What is left is skills, and if those are unstructured then the workflow was never deterministic to begin with. His warning is that ungoverned skills become a recognizable kind of technical debt. Teams on the same stack quietly build the same skill repeatedly. Quality decays unless skills are retested against new models, not just against their original task. Nobody knows who owns one. Composability does not happen by accident. There is a real security edge too, since skills carry scripts, so pulling a public one without a checking pipeline is a supply chain decision. His fix borrows from the microservices era wholesale: a catalog with metadata and search, dependencies, versioning so a harness pulls the current release, access control, and named human owners across architecture, infrastructure and security. He is blunt that governance is where technology stops solving it. Speaker info: / imad-touil Timestamps: 0:00 - Three questions, and the hands that drop 1:33 - The agentic stack, inner loop and outer 2:56 - Why specify, plan, task, implement is one step 4:19 - What the full lifecycle actually contains 5:39 - Organizations run many lifecycles, not one 7:03 - Where the know how actually lives 8:27 - Designing skills like microservices 9:48 - Progressive disclosure and token cost 11:09 - Composing regulation skills into an audit trail 12:30 - Ungoverned skills as technical debt 13:52 - What a central skills platform needs 15:14 - Governance, and who owns which domain 16:34 - Simulating fifteen teams over six months 17:58 - Applying the same idea to whole workflows 19:22 - Registries, evaluation and auto evolving skills