{
  "video": {
    "id": "M05vON8i0aI",
    "title": "AI-Native Organisations Run on Skills: How to Structure and Scale Them — Imad Touil, QuantumBlack",
    "duration": 1231,
    "upload_date": null,
    "channel": "AI Engineer",
    "source": "AI Engineer"
  },
  "analysis": {
    "video_id": "M05vON8i0aI",
    "title": "AI-Native Organisations Run on Skills: How to Structure and Scale Them — Imad Touil, QuantumBlack",
    "one_liner": "A QuantumBlack distinguished engineer argues that skills are where an organisation's know-how actually lives in the agentic stack, and that without a governed, centralised skills registry — catalog, versioning, ownership, access control, evals — teams generate a new class of technical debt: duplication, degrading quality and rising token cost.",
    "summary": "Imad Touil maps the agentic software stack — an inner code-agent harness loop (context manager, tools/MCPs, memory and state, skills loader) and an outer workflow loop (skills, sub-agents, MCP servers, hooks) sitting on enablement components like an MCP gateway, model gateway, knowledge graph and workflow marketplace. He argues that of the four workflow components, hooks, MCP servers and sub-agents are largely given or borrowed, so all real organisational know-how ends up in skills — and skills are what make workflows deterministic. He then treats skill design as the microservices problem again (reusable, modular, discoverable, portable, specialised, composable, consistent, cost-efficient) and shows a six-month simulation of 15 teams to contrast ungoverned skill sprawl with a centralised, governed catalog.",
    "key_points": [
      "Opening show of hands: many in the room create and use skills, fewer share them within teams, and only a few hands stay up for governing and maintaining skills across the organisation — that gap is the whole talk.",
      "The agentic stack has two loops: the inner code-agent harness (context manager, tools/MCPs, memory and state, skills loader) and the outer workflow loop (skills, sub-agents, MCP servers, hooks), with enablement layers — environment sandbox, MCP gateway, model gateway, a knowledge graph abstracting IT core systems/codebase/skills registry, and a workflow marketplace.",
      "The familiar specify → design/plan → tasks → implement loop that coding agents are shaped around is only one step — a product increment — inside a real enterprise lifecycle that also spans product strategy, market/competitive research and customer interviews, discovery, data preparation and data-product delivery, platform engineering ops, launch, and performance/incident optimisation; and one organisation runs many different SDLCs at once (mobile, internal platform, customer-facing).",
      "Of the four workflow components, hooks just fire on events, sub-agents exist mainly to minimise the context window, and almost nobody actually builds MCPs — they consume ones their vendors ship — so 'all of your knowhow is actually at the skills level'. Workflows are 'harness blueprints' that shape harness behaviour at runtime.",
      "Timeline of adoption: Anthropic published the first skills article about eight months ago, an open standard followed two months later, and by around February most agent harnesses had adopted it — often invisibly, visible only when you watch the agent pull skills as it works. A snapshot of public GitHub repos and skills registries shows creation rising sharply.",
      "A skills-bench comparison of the latest models on the same tasks (auto engineering and cyber security) showed models do well without skills — and clearly higher with skills applied, because skills make the outcome more deterministic.",
      "Skill design principles are lifted from microservices: reusable, modular, discoverable, portable across workflows and harnesses (a skill written for Claude Code moves to Cursor and 'it's going to just work'), specialised to one task rather than a monolith, composable without conflicting duplication, consistent/deterministic, and cost-efficient — progressive disclosure puts the right amount of skill in context at the right time and cuts token usage.",
      "Worked example: a composable regulation skill set — data retention policy, disclosure standards, GDPR rules, fill-in templates — pulled automatically at runtime by a regulatory disclosure review workflow, producing a deterministic audit report and improvements that loop back into the codebase.",
      "Ungoverned skills create a new class of technical debt across seven axes: duplication (teams on the same stack rebuild the same skills), quality decay (skills must be validated against new models, not just tasks), discoverability, ownership (the Backstage/IDP service-catalog problem again), composability (needs a domain-driven approach), security (public skills can carry prompt injection and skills execute scripts), and permissions (some skills encode sensitive business logic).",
      "A simulation of 15 teams of 5–12 people over six months tracked skills per engineer, average skills pulled per day, cross-team duplication ratio, and quality/security ratio: ungoverned, maturity varies wildly team to team, one team shows low-medium productivity with high cost, and missing skills mean engineers vibe-code back and forth to steer the agent, burning tokens and time. Governed, teams converge on common ground and the harness pulls an existing published skill instead of building a new one."
    ],
    "takeaways": [
      "Stand up a centralised skills platform, not just individual skill files: a catalog with metadata that is searchable, an MCP plugged into that catalog, a CLI to pull skills into your IDE or sandbox, plus dependency tracking, versioning and lifecycle (so an agent can detect and pull the latest version mid-task), access control, evaluation and observability.",
      "Assign human governance owners — architects, engineering leads, infra leads, cyber leads — to specific skill domains, because at the governance step 'technology stops solving the problem'.",
      "Design skills like microservices: one specialised task per skill, composable rather than monolithic, portable across harnesses, and structured for progressive disclosure to cut context cost.",
      "Put a security pipeline in front of skills you pull from public sources — skills carry executable scripts and can carry prompt injection — and gate sensitive business-logic skills with access control.",
      "Evaluate skills statically against Anthropic's skill best practices as a starting point (badly structured or badly invoked skills are probably low quality), and treat governance as the prerequisite before turning on any auto-evolving closed loop."
    ],
    "topics": [
      "agent-skills",
      "agentic-workflows",
      "governance",
      "platform-engineering",
      "mcp",
      "evals",
      "context-engineering",
      "enterprise-ai"
    ],
    "tools": [
      "QuantumBlack",
      "Anthropic",
      "Claude Code",
      "Cursor",
      "MCP (Model Context Protocol)",
      "Backstage",
      "GitHub",
      "GDPR"
    ],
    "quotes": [
      {
        "text": "So at the end of the day you will find all of your knowhow is actually at the skills level.",
        "at": "06:42",
        "url": "https://www.youtube.com/watch?v=M05vON8i0aI&t=402s"
      },
      {
        "text": "This define a new unit right that makes your knowhow in your organization executable portable and cheap.",
        "at": "09:51",
        "url": "https://www.youtube.com/watch?v=M05vON8i0aI&t=591s"
      },
      {
        "text": "all of this is actually play around a governance and this is where technology stop solving the problem right",
        "at": "14:47",
        "url": "https://www.youtube.com/watch?v=M05vON8i0aI&t=887s"
      },
      {
        "text": "The moment you govern, you publish one skill, the next engineer trying to build a new skill, the coding agent harness will identify this skill that is already available and pull it.",
        "at": "17:36",
        "url": "https://www.youtube.com/watch?v=M05vON8i0aI&t=1056s"
      }
    ],
    "words": 4463
  },
  "summary_url": "/#M05vON8i0aI",
  "transcript": {
    "html": "/transcripts/M05vON8i0aI.html",
    "txt": "/transcripts/M05vON8i0aI.txt",
    "vtt": "/transcripts/M05vON8i0aI.vtt"
  }
}