Code of Leadership #86 Materials: AI Agents and Developer Productivity
The recording and summary of Code of Leadership episode 86 are ready. The episode went out on October 22, 2026, under the title “Agents Write the Code. How Do We Know Development Has Become More Efficient?” We talked with Vladimir Dolgov, the creator of DevOptify, who previously led the AI transformation of software development at Sber.
There's more code, PRs are closed more often, and tokens are being used. What is the business getting out of it? We built the conversation around that question.
What we discussed
🔸 What to count as a result
Vladimir described how Sber moved from the share of AI-written code to counting PRs, then completed tasks. But you can always slice tasks into smaller pieces. You also need to account for work size, errors, and work sent back for revision.
🔸 How a good-looking chart can mislead
I gave the example of a team that built a platform and then used it for a series of similar product changes. The number of closed tasks rose, but the mix of work had changed as well. That matters before you declare victory.
🔸 Why agents make us relearn planning
Vladimir's team initially allowed too much time for tasks, then swung the other way and began underestimating them. In the DevOptify demo, we looked at how that shows up in the data alongside parallel tasks and overtime.
🔸 Where humans still do the work
An agent needs context, tools, and limits on its authority. The volume of generated code is constrained by review capacity. Vladimir described the sampling his team uses for human review. Applying that practice to every product would require some qualifications.
🔸 What model spending tells us
In one example a large model was used for a bulk rename that an IDE command could have handled. High cost alone doesn't prove poor work. It depends on the task.
🔸 How to maintain trust in analytics
Work sent back for revision and late-night work give you a reason to talk with an engineer. We discussed giving employees access to the same data and a way to challenge a manager's conclusions, for instance by pointing out unplanned work that threw off the original schedule.
One qualification about the demo: at the time of the conversation, the results of large DevOptify pilots were still pending. Vladimir described the evaluation cycle: collect baseline data, change how work is done, and see what happens afterwards.
Episode materials
📌 Episode page with timestamps
🎬 Video: YouTube and VK Video
🎧 Audio: Podster and Apple Podcasts
📖 Written summary
What changes in your team tell you that agents are actually helping? I'm especially interested in cases where the metrics and your impressions diverge.
#CodeOfLeadership #AI4SDLC #Agents #Management #Metrics #DevTools