Decoding the Research
In June 2026, Anthropic published a paper analyzing approximately 400,000 Claude Code sessions from over 235,000 users. The study, which took place between October 2025 and April 2026, aimed to understand who succeeds with AI coding agents and why. The core
finding was a surprise: success was more closely tied to a user's domain expertise—their deep understanding of the problem they were trying to solve—than their formal software engineering background. While software engineers had slightly higher success rates, professionals in fields like management, law, and sales performed nearly as well, with all major occupations falling within a few percentage points of each other. This suggests the value of these tools extends far beyond traditional coding roles.
The New Division of Labor
The research highlights a clear division of labor in human-AI collaboration. Users are responsible for about 70% of the planning decisions—defining the 'what'—while the AI handles roughly 80% of the execution—the 'how'. This dynamic is where domain expertise becomes critical. An expert user who deeply understands the business logic, the customer needs, or the specific industry problem can provide clearer, more effective prompts. As a result, the AI performs more actions and generates significantly more useful output. The study found that expert users prompted Claude to perform 2.4 times more actions and generate over five times more output per instruction compared to novices. This shift moves the bottleneck from typing code to defining the problem with precision.
Rethinking Project Planning
This new reality demands a change in how technical teams approach project planning. If an AI agent can handle much of the implementation, the most valuable human contribution becomes high-level strategic thinking. Planning sessions must now focus more on crafting the perfect problem statement and providing the AI with the right context. Senior engineers' roles may evolve from primarily writing complex code to architecting systems and guiding AI agents. Their expertise is needed to validate the AI's proposed plan, catch architectural flaws before implementation begins, and ensure the AI's output aligns with the broader goals of the project. Success is no longer just about building things right, but about ensuring the team is building the right thing from the outset.
Supercharging the Code Review
The code review process is also being transformed. As AI assistants accelerate code generation, the volume of code needing review has become a significant bottleneck. Systems like Claude's multi-agent code review aim to solve this by deploying several AI agents that analyze pull requests in parallel, checking for logic errors, security flaws, and performance issues. One internal Anthropic case study found that meaningful review rates for pull requests jumped from 16% to 54% after implementing AI reviewers. However, this doesn't make human oversight obsolete. Instead, it elevates the human's role to focus on more substantive issues like architectural integrity and alignment with business goals, leaving the AI to catch syntax errors and common bugs.
The Enduring Value of Domain Knowledge
Perhaps the most important takeaway is the rising value of domain knowledge. An AI assistant paired with a developer who lacks business context can produce code that is technically correct but functionally wrong. The most valuable developers in an AI-driven world are those who understand the 'why' behind the code. They can ask better questions, spot subtle errors in AI-generated logic, and guide the tool toward a solution that truly solves the business problem. The research showed that while time spent debugging code fell from 33% to 19% over the study period, the overall value of the tasks being performed increased by 27%. This indicates a shift toward more complex, higher-value work, where deep expertise is indispensable.
