← back
Tribal Dungeons of Global Shipping: AI Agents at Global Scale — Dmitry Buykin, Maersk
Takeaway
Reliable enterprise agents emerge from continuously refining executable operational knowledge and bounded workflows with domain experts.
Summary
- Maersk’s production agents handle shipping exceptions across interacting workflows and legacy systems whose operational knowledge is often buried in screenshot-based SOPs.
- Agent-ready procedures encode preconditions, identifiers, decisions, backend actions, validation, recovery, and evidence of success.
- The speaker reports more than 200 production instances and over 100,000 corrections across nine months, with expert review and failure clustering directing improvements.
- Real cases are replayed with production write capabilities disabled; critical paths retain approval gates, and successful sequences become reusable composite tools.
maerskenterprise-agentsoperational-knowledge
Original description
Maersk's standard operating procedures were screenshots. A sequence of images showing what a person sees and where they click, which is a perfectly good record for a human and useless to an agent. Dmitry Buykin calls the gap tribal dungeons: the knowledge exists, just not in a form anything can execute safely. An agent version of the same procedure needs preconditions, decisions, identifiers, backend calls, validation, recovery and evidence that it actually worked. Most of the project was that translation, negotiated with the people who own the process, because experts own the what and agents own the how. His sharpest point is about where the engineering actually lives. The agent loop is not the system. The refining loop around it is, and the corpus of procedures outweighs the runtime roughly twenty to one, because the same shipping step means different things in different countries. Accuracy was not designed up front in a diagram, it was earned through more than 100,000 corrections over nine months, with heat maps turning traces into priorities and a single cell often costing the team a month or two. A correction only counts once it becomes an executable change, which is the line between an opinion and a production fix. Discovery needs agent freedom, production needs a cage, and a harness exists to make the dumb mistakes impossible rather than to give the model more room. Speaker info: https://x.com/tzakus / buykin Timestamps: 0:00 - The long tail is the expensive part 2:04 - What an SOP has to become for an agent 3:56 - The refining loop is the system 4:50 - Running 200 instances against legacy backends 5:47 - Triage, traces, and shared evidence 6:44 - Where vibe coding and spec driven work run out 7:39 - A hundred thousand corrections, and heat maps 8:36 - Please be careful is not a guardrail 9:31 - Five moves, and compounding improvement 10:31 - Composite tools, and why they skip MCP