At Maersk, most of one agent system is prose, and the skill is deciding what survives

Dmitry Buykin of Maersk argued on the AI Engineer show that in his agent system most of the "code" is plain-language instructions. Writing those instructions is no longer the hard part. The hard part is deciding which behaviour must survive each change. Buykin said business experts should own the decisions, and the model should only choose among them.

What was said

Buykin said his team runs the system across 110 countries, with roughly 10,000 scenarios written as prose. By his estimate, about three quarters of the system is prose.

That raised a testing problem. His team borrowed mutation testing, where you break one sentence on purpose and watch for a failing test. Almost no test failed. "Prose has no compiler, so we build one," he said.

He also argued that rewriting is no longer the obstacle. "With agents, a rewrite is easy. The hard part is knowing what has to survive it." Tests can answer that question for ordinary code, he said. For prose, only people can.

His team's answer is a split of roles. Experts hold the decision table. The model's job is limited to choosing one row from it. When nothing fits, a human steps in. Buykin said a repeated path becomes code only once it is the same across countries. Anything that still varies stays as prose.

Experts cannot see how systems connect underneath. So developers dig through traces and logs and write those links down, so the runtime can check them. Experts review each task exactly as the agent saw it.

Why it matters

Our reading: this changes what skilled work means. The scarce skills are knowing which rules are the business and which are noise, and being able to say so with authority. Typing the instructions is the easy part.

For managers who buy or run agent systems, this suggests some questions. Who owns the decision table? Who signs the prose? How are instructions tested when ordinary tests do not notice a changed sentence?

It also suggests a training path. Developers become the people who make business rules checkable. Experts become reviewers of exactly what the agent saw, not distant approvers. Buykin noted that people tire, so his team spends their judgment only where it is needed.

The other side

Buykin did not present this as settled. He said the boundary between code and prose moves both ways. His first workflow tools made things worse. His team also shifted a few hundred lines of logic from code into prose, because experts could not read it.

He also admitted a limit of expert ownership. The system picks up its corrector's habits: "Conway said the system copies the org chart. Ours copies whoever corrects it." He said he found his own principles and biases built into the design.

He said the right answer depends on a group of experts, because no single person knows it. This is one company's account of its own system. The excerpts give no independent measurement of how well it works, or whether it suits smaller teams without 110 countries of experts.