Dmitry Buykin of Maersk walks through an agent-maintained system spanning country-specific business processes. This 18-minute AI Engineer talk uses a failed wholesale rewrite to motivate smaller, evidence-led maintenance loops.

Why the rewrite failed

  • Agents kept adding code, tests and prose instructions; a purge removed useful behavior, and the sprawl grew back.
  • His alternative: keep change below what people can verify through weekly sweeps and feedback loops.
  • An order-amount change illustrates the stakes: the same rule can be valid in one country and unsafe in another.

Where decisions live

  • Experts specify business rules; developers use traces to connect those rules to systems. An expert reviews the task as the runtime agent saw it.
  • Repeated, cross-country instruction paths become workflow code; variable judgments stay in an expert-readable decision table. A missing row goes to a person.
  • Lint the instructions, test them by intentionally changing a sentence, and watch for contradictions between code and prose.

Keep the runtime bounded

  • The runtime selects the few facts a model needs. Loading the entire order led the model to invent an unrelated pricing check.
  • The model passes a handle to a record the runtime holds; the runtime validates it before writes. Permission and context limits both matter.
  • A journal makes retried writes happen once; guards check the action before and after it lands.

Verify and prune continuously

  • Replay real tasks against old and new versions. A refactor once lost the above-limit branch that required a team lead.
  • A second agent checks with a checklist, without the author’s reasoning; unresolved work stays visible in a ledger.
  • Expert corrections become evaluations across related country slices, but fixes land only where the same failure is confirmed.
  • Buykin describes ratchets and weekly retirement of unreachable variants, including the tests that kept dead variants alive.

Caveat: The scale, results and claimed guardrails are Buykin’s account of his team’s system, not an independent audit.

“Verification speed is the limit.”