{
  "video": {
    "id": "dQ-_i1tZiws",
    "title": "Tribal Dungeons of Global Shipping: AI Agents at Global Scale — Dmitry Buykin, Maersk",
    "duration": 722,
    "upload_date": null,
    "channel": "AI Engineer",
    "source": "AI Engineer"
  },
  "analysis": {
    "video_id": "dQ-_i1tZiws",
    "title": "Tribal Dungeons of Global Shipping: AI Agents at Global Scale — Dmitry Buykin, Maersk",
    "one_liner": "A Maersk practitioner report on running 200+ agent instances in global shipping ops, arguing that the agent loop is not the system — the refining loop around it is, built from an SOP corpus 20x bigger than the runtime and over 100,000 expert corrections in 9 months.",
    "summary": "Buykin describes production AI agents supporting Maersk's global shipping operations, where the easy majority of work is already automated and what remains is a long tail of exceptions spanning many incomplete legacy systems. He calls the core problem \"tribal dungeons\": the operational knowledge exists but not in a form an agent can execute — legacy SOPs are sequences of screenshots showing what a person sees and clicks, while an agent SOP needs preconditions, decisions, identifiers, back-end calls, validation, recovery and evidence of execution. The architecture is three parts — SOP memory (the corpus), execution runtime, and feedback capture — with quality earned one correction at a time through replayed real examples with writes disabled, trace review shared between experts and engineers, and heat maps that prioritise where the team works. The closing position: the real outcome was the methodology, not the agent, and they deliberately don't use MCP.",
    "key_points": [
      "On paper a shipment is one workflow; in reality it is an orchestration of many parallel state machines, and the moment one drifts you get exception work — the expensive long tail that exceeds what the systems were built to handle.",
      "\"Tribal dungeons\": the knowledge exists but not in an executable form, so you cannot safely run a process the organisation cannot represent. Legacy SOPs are screenshots in sequence — an agent SOP needs preconditions, decisions, identifiers, back-end calls, validation, recovery and evidence of successful execution.",
      "Three-part architecture: SOP corpus (memory), execution runtime, and feedback capture. The corpus is the company's process memory, aligned per country, and is roughly 20:1 bigger than the runtime.",
      "Production scale today: over 200 concurrent instances; latency ranges from a few minutes up to 10 minutes, bounded mainly by legacy back ends that cannot go faster than the agent loop.",
      "Over 100,000 corrections accumulated over the last 9 months; accuracy was not designed up front in one diagram but earned one small correction at a time.",
      "Expert time is the bottleneck, so a bench clusters failures into actionable triage; traces are the shared evidence experts and engineers review together; a correction only counts when it becomes an executable change.",
      "Quality comes from replaying real examples with write access disabled (to protect production systems) and checking whether behaviour improved — \"not from vibes, not from a bigger model.\"",
      "Failure-to-fix mapping: wrong workflow → classifier eval; wrong write → write gate; wrong assumption → review. Preventive measures eliminate the unsafe path rather than merely warning; review and approval stay in the loop on critical paths.",
      "They deliberately do not use MCP — the systems are bloated, so they distil responses and tune tools through function calling to control quality.",
      "Repeatable successful step sequences are merged into bigger composite tools and reusable snippets, which can then be rolled out to hundreds of countries in one go."
    ],
    "takeaways": [
      "Convert screenshot-based SOPs into agent-executable specs with explicit preconditions, decisions, identifiers, back-end calls, validation, recovery and success evidence — and treat that corpus, not the agent loop, as the asset.",
      "Split ownership: experts own the what, agents own the how, and every exception becomes a guardrail. Budget most of the effort for the translation and negotiation between the two.",
      "Build the replay harness: re-run real production examples with writes disabled and measure whether behaviour improved, instead of reaching for a bigger model.",
      "Make corrections executable — cluster failures, review shared traces with experts, use heat maps to prioritise (one red cell took the whole team of engineers and agents about one to two months), and treat an agent failure as where investigation starts.",
      "Match the preventive control to the failure class (classifier eval / write gate / review), and give discovery agent freedom while production gets a cage that makes dumb mistakes impossible.",
      "Consider skipping MCP for bloated legacy back ends and hand-tuning function-calling tools instead, so you control response size and quality."
    ],
    "topics": [
      "agents",
      "production-ops",
      "sops",
      "evals",
      "guardrails",
      "observability",
      "legacy-integration",
      "tool-calling"
    ],
    "tools": [
      "Maersk",
      "MCP (Model Context Protocol) — explicitly not used",
      "function calling",
      "SAP (mentioned as a slide mix-up)"
    ],
    "quotes": [
      {
        "text": "The agent loop is not the system. The refining loop around the agent is the system",
        "at": "03:48",
        "url": "https://www.youtube.com/watch?v=dQ-_i1tZiws&t=228s"
      },
      {
        "text": "Experts own the what, agents own the how. And exception becomes a guardrail.",
        "at": "03:21",
        "url": "https://www.youtube.com/watch?v=dQ-_i1tZiws&t=201s"
      },
      {
        "text": "discovery needs agent freedom and production needs a cage. Uh the harness isn't there to give the agent more room. It's there to make the dumb mistakes impossible.",
        "at": "08:22",
        "url": "https://www.youtube.com/watch?v=dQ-_i1tZiws&t=502s"
      },
      {
        "text": "we're not using MCPS because uh for us it's uh always not the best choice. So because all all systems usually really bloated and we have to distill responses and uh tune the tools through function calling",
        "at": "10:57",
        "url": "https://www.youtube.com/watch?v=dQ-_i1tZiws&t=657s"
      }
    ],
    "words": 1570
  },
  "summary_url": "/#dQ-_i1tZiws",
  "transcript": {
    "html": "/transcripts/dQ-_i1tZiws.html",
    "txt": "/transcripts/dQ-_i1tZiws.txt",
    "vtt": "/transcripts/dQ-_i1tZiws.vtt"
  }
}