Buykin describes global shipping as a system where the easy paths are already automated and the costly work lives in unusual exceptions. Human standard operating procedures often contain screenshots and assumptions that an experienced employee can interpret, but an agent needs every precondition, identifier, backend call, validation step and recovery path made explicit before it can act safely.
Maersk organizes this work around three connected components: a corpus of executable procedures, a bounded agent runtimeAn AI agent runtime is the execution system that manages an agent's model calls, tools, state, permissions, and task lifecycle. and a bench where subject-matter experts inspect traces and correct behavior. More than 200 production instances operate across legacy systems that can take minutes to respond, so the surrounding architecture must preserve state, capture evidence and make failures understandableAI agent observability makes an agent's state, actions, tool use, failures, resource use, and outcomes visible enough to understand and operate it. rather than treating the model as the whole product.
The limiting resource is expert attention. Failure clustering and heatmaps help teams direct that attention toward repeated patterns, while shared traces give engineers and operators the same evidence. A correction only compounds when it becomes an executable change to a procedure, guardrailA safeguard is a measure designed to prevent, limit, detect, or recover from unwanted behavior or harm in an AI system. or evaluation case instead of remaining a comment that the next run can ignore.
Buykin recommends allowing freedom during discovery but putting production behavior inside a cage. Critical approvals remain humanHuman-in-the-loop AI keeps a person involved in reviewing, correcting, approving, or guiding an AI system's work., permissions are disabled during replay testsAn agent permission boundary limits the information, tools and actions an AI agent can use during a task., and every known failure mode should map to a preventive measure that makes the unsafe path structurally impossible. The resulting quality comes from replaying real cases and learning from roughly 100,000 corrections, not simply swapping in a bigger model. Closing Q&A and invitations are omitted.
Watch on YouTube
