Agents Write Code. Is Development More Effective?

Vladimir Dolgov describes moving from corporate developer tools to his own company and DevOptify. Managers have to revisit assumptions about task complexity and delivery time as agents change engineering work. Vladimir uses Cursor, Codex, and Claude; Alexander discusses combining tools and keeping instructions portable between them.

The measurement story at Sber moves from the share of AI-generated code to PR counts and then to tasks recognized as outcomes by the business. Vladimir argues for several levels of metrics and countermetrics: code volume needs defect and rework checks, while task counts need task-size context. Alexander describes a throughput increase caused by a shift from platform development to repeatable product changes. Product lifecycle stages also change what metrics mean.

Agents need knowledge, tools, a working environment, and limits on their authority. Their output can exceed human review capacity; Vladimir describes selective human review as his own team’s practice. The DevOptify demonstration covers planning accuracy, model use, parallel tasks, and overtime. An experiment with changes in prompt tone is presented as a reason to investigate a situation, rather than a reliable diagnosis.

Practical analytics questions include which tasks reopen, how much time fixes consume, and why someone is working at night. Events from IDEs, browsers, and CLIs are grouped into work sessions; time in a tracker status is different from time spent doing the work. They discuss cloud and on-premises deployment, MCP, limiting collection to work projects, and employees’ ability to challenge an assessment. Vladimir describes pilots as baseline measurement, a management intervention, and further observation; the large pilots have yet to produce results at the time of the conversation.