Anthropic study finds domain knowledge beats raw coding skill

For the past two years, the consensus was simple: to get the most out of an AI coding agent, you probably needed to be a software engineer.

Anthropic just proved that assumption dead wrong.

Anthropic released findings from an extensive study analyzing roughly 400,000 Claude Code sessions across 235,000 users between October 2025 and April 2026. It is one of the most comprehensive real-world looks at how people actually use autonomous coding tools.

The headline takeaway? Success with a coding agent is driven almost entirely by your domain expertise, not your programming skills.

Humans plan, Claude executes

The usage patterns reveal a clean, sharp division of labor.

In a typical session, humans make around 70% of the planning decisions, defining what needs to be done. Claude handles roughly 80% of the execution decisions, figuring out how to actually implement it.

A single user prompt triggers an average of 10 Claude actions and produces 2,400 words of output. Seasoned users command an average of 12 actions and 3,200 words per prompt, compared to novices who trigger just 5 actions and 600 words.

Here is the twist: "expert" does not mean software engineer. The study found that a lawyer who understands exactly which clauses a compliance script needs to flag succeeds at getting working code almost as often as a trained developer.

Across nearly every major occupation, people with deep domain context achieved their technical goals with verifiable evidence, like passing tests or committed code, at rates comparable to professional engineers.

Beyond debugging to end-to-end tasks

The way people interact with Claude Code changed dramatically over those seven months.

The share of sessions spent merely hunting down bugs dropped by nearly half. Instead, users shifted toward end-to-end agentic workflows: deploying software, configuring pipelines, analyzing complex data, and generating non-code documentation.

Looking across all sessions, 56% still revolve directly around code, split between building new features (25%), fixing bugs (26%), and testing or orchestrating pipelines (5%). Software operations account for 17%, while planning and exploring architecture take up 14%.

As users trusted the tool with larger scopes, the estimated freelance value of typical tasks jumped by roughly 25%. GitHub projects with coding agent activity have more than doubled since late 2025, and Claude Code users now log an average of 20 hours per week inside the tool.

Why it matters

We think this study completely shifts the labor math around software.

For decades, the moat in tech was knowing how to write syntax and navigate frameworks. Claude Code flips that on its head. The syntax layer is being absorbed by the machine, meaning the ultimate leverage belongs to whoever truly understands the problem.

For domain experts—accountants, doctors, operations managers, and legal pros—the barrier to building custom software has essentially vanished. If you know how the business machinery works, you can direct an agent to build the tooling yourself.

For junior developers whose day job consists mostly of translating Jira tickets into standard boilerplate, this is an existential warning. The pure implementation layer is rapidly eroding.

Anthropic played this smartly. Instead of hyping total human obsolescence, they backed up the conversation with real telemetry: the winning formula is a domain expert acting as the architect, with the AI serving as the tireless junior engineer.