An Anthropic study, based on the analysis of 400,000 Claude Code sessions, showed that the effectiveness of using AI agents directly depends on the user's level of domain expertise, rather than their syntax writing skills.
What Happened
According to Anthropic data for the period from October 2025 to April 2026, experts using Claude Code receive 2.4 times more actions from the AI (averaging 12 compared to 5 for beginners) and 5 times more text output per turn (~3200 vs ~600 words). A shift in task structure was also recorded: the share of sessions related to bug fixing decreased from 33% to 19%, giving way to data analysis and software management.
Context
In agentic scenarios, a clear division of roles is observed: humans make about 70% of the decisions regarding what exactly needs to be done (intent), while the AI agent takes on about 80% of the decisions regarding implementation. This transforms programming from a process of writing lines into a process of managing intentions.
Why It Matters for the Industry
For the IT industry, this means a shift in value from writing syntax to system design and task specification. The developer competency profile is transforming from a classic "coder" to a "conductor" of AI agents, which will require a revision of educational standards in Computer Science and new approaches to model performance evaluations (evals).
Why It Matters for Users
For users, this means that to work effectively with AI agents, it is not necessary to be a professional developer, but it is critically important to deeply understand the essence of the delegated task. The higher your domain expertise, the more powerful your "leverage" becomes when managing the neural network.
What Is Not Yet Known / Limitations
Expert opinions regarding the consequences vary: from predictions of productivity growth to concerns about legal and confidentiality risks.
Sources
Author
Look at AI, Editorial Team