Summary
Dmitry Buykin delivers a practitioner report from Maersk on deploying AI agents to handle exception management in global shipping operations — the messy long tail of shipments that break from the happy path and require expert orchestration across multiple legacy systems. The talk skips the conceptual pitch and goes directly to what the team calls ‘tribal dungeons’: operational knowledge that exists in human heads and screenshot-based SOPs but is not in a form an agent can safely execute.
The core engineering challenge is translating those legacy SOPs into agent-executable procedures with explicit preconditions, decision trees, backend calls, validation steps, and evidence of successful execution. This translation requires ongoing negotiation between domain experts (who own the ‘what’) and engineers (who own the ‘how’). The resulting system runs over 200 concurrent agent instances with latencies ranging from a few minutes to ten minutes — bounded by the speed of the legacy systems being orchestrated, not the agent loop itself.
What makes the talk quantitatively notable is the scale of the correction loop: over 100,000 corrections accumulated across nine months of development, with heat maps clustering agent failures into prioritized work items that engineers and experts review together. Buykin argues that accuracy at this scale is not designed upfront but earned one correction at a time, and that a robust harness around the agent — classifier evals, write gates, assumption review, human-in-the-loop on critical paths — is what separates a reliable production system from a demo.
📺 Source: AI Engineer · Published August 29, 2026
🏷️ Format: Workflow Case Study
