Research Insights Made Simple 20Why coding agents do not cancel the examination (Category AI4SDLC)

What happens when a coding agent selects files, runs commands, and writes an implementation? Thursday, 16 July, with 16:00 before 17:00 MSC discuss Evgeniy Sergeev in "Research Insights Made Simple - Season" 1, Episode 20" Whitepaper Anthropic: "Agentic coding and persistent returns to expertise"

Eugene Sergeev Director of Engineering at Flo Health. He develops product engineering teams and is engaged in the practical implementation of AI in development: coding agents, eval-driven development, quality gates and AI-assisted workflows. Zhenyu is interested in how to make agent work manageable and verifiable — and what happens to engineering expertise when code becomes cheaper to write and the cost of making bad decisions remains high.

If we go back to the papyra, the authors analyzed the 398 198 interactive sessions 234 751 Claude Code users concluded that user expertise in a task is associated with deeper delegation to an agent and more frequent session success. But this is an observational sample of a single product, not an industry-wide performance experiment. So let's not just talk about the numbers, but the limits of what they prove: Why expertise here is knowledge of a specific task, and not grade or experience;

• What does the division of labor mean? 70% of decisions that what do, but only 20% about how; Why expert sessions run about twice as many agents, but that doesn’t mean twice as much productivity. Is it possible to count growth as “verified success”? 14,5percentage 32,9% effect of examination or we see only correlation; Where is the boundary between a successful session, accepted PR, reliable production code and business benefits
• How teams measure verification time, rework, defects, and knowledge retention, not just generation rate.

Separately, let’s talk about the organizational paradox: an experienced engineer is able to disperse an agent more strongly, but if all the execution is given to the machine, where will the next experienced engineers come from?

For me, this isn't about whether Claude Code will replace a developer. The question is more practical: what kind of expertise becomes scarce when the code becomes cheaper, and the cost of a wrong decision remains with the team. Come on. live Thursday, 16 July, with 16:00 before 17:00 MSK. We will analyze the methodology, argue with the conclusions and translate the whitepaper into questions to the real AI4SDLC.

P.S. The whitepaper itself I have already examined in two parts: 1 and 2.

#AI #AI4SDLC #Research #Engineering #Agents #Metrics #DevEx