Agents Made Code Free. Maintaining It Didn't Get Cheaper. - Dmitry Buykin, Maersk
When agents write most of a production codebase, the bottleneck shifts to review and maintenance. Buykin describes weekly evidence loops for keeping changes understandable and safe to ship.
Buykin’s system runs across **110 countries** and about **10,000 scenarios**. It keeps expert-owned rules readable, gives agents only task-relevant context, and uses runtime checks to constrain changes. A replay of a proposed check against thousands of past sessions would have triggered on less than **1%**.
For coding-agent workflows, make repeated agent paths explicit tools, preserve human judgment where rules vary, and connect user feedback to evaluations. Keep durable memory in files and require evidence such as replaying real tasks; a second agent can check changes against a checklist without seeing the author’s reasoning.
Buykin’s system runs across **110 countries** and about **10,000 scenarios**. It keeps expert-owned rules readable, gives agents only task-relevant context, and uses runtime checks to constrain changes. A replay of a proposed check against thousands of past sessions would have triggered on less than **1%**. For coding-agent workflows, make repeated agent paths explicit tools, preserve human judgment where rules vary, and connect user feedback to evaluations. Keep durable memory in files and require evidence such as replaying real tasks; a second agent can check changes against a checklist without seeing the author’s reasoning. The approach depends on people reviewing the decisions agents cannot reliably infer, and on having real tasks to replay. The talk describes one production system; it does not establish that the same checks or thresholds fit every codebase.
This account emphasizes how production reliability depends on keeping expert rules readable, limiting agent context, and replaying proposed checks against historical sessions. The reported under-1% trigger rate illustrates how replay can narrow a check before deployment, without establishing that the threshold suits other systems.