The narrative around AI coding tools has been simple: democratise software. Anyone can build now. And there's some truth in it. But Anthropic just published research based on around 400,000 Claude Code sessions that tells a more interesting story.
Yes, non-coders can ship. But expertise still compounds.
What the research actually found
Anthropic looked at sessions from October 2025 through April 2026, tracking how different types of users work with Claude Code. The headline numbers are striking. Expert users trigger around 12 Claude actions per prompt and get roughly 3,200 words of output. Novices? About 5 actions and 600 words. More than double the activity, five times the output, from the same tool on the same task.
On success rates, the gap is equally clear. Verified success in novice sessions sits at 15%, with 77% reaching partial success. Intermediate to expert sessions hit 28 to 33% verified success and 91 to 92% partial. When a session runs into trouble, novices abandon it 19% of the time. Experts, just 5 to 7%.
The division of labour is interesting too. Users make around 70% of planning decisions (what to build, what the goal is), while Claude handles about 80% of execution decisions (how to do it, which files to edit, which approach to take). A typical session is four prompts from the human, each triggering about ten Claude actions. Sometimes more than a hundred.
Why this matters
Here's the thing that reframes the whole "anyone can code now" conversation: success isn't strongly tied to whether you're a software engineer. Non-software professionals hit 29% verified success compared to 34% for software folks. A five percentage point gap, across completely different professional backgrounds. Management, sales, and legal are among the fastest-growing groups using the tool.
What actually drives outcomes isn't coding background. It's domain expertise. Knowing what good looks like. Knowing what the software is supposed to do. Knowing when the agent has misunderstood the goal. The research puts it directly: "The more domain expertise a person brings to a session, the more work Claude does per instruction."
That's a significant finding. The limiting factor in agentic coding isn't your ability to write a loop or debug a function. It's your ability to direct, evaluate, and course-correct. The people who are best at that are the ones who understand the problem space deeply, not necessarily the people who've spent years writing code.
The economic value of typical tasks rose 27% over the seven months of the study, with building tasks up 43%, operating tasks up 34%, and fixing tasks up 32%. More value, coming faster, from a broader set of people.
How Dakik can help you put this to work
We build products with these tools every day, and this research matches what we see in practice. The teams shipping fastest aren't necessarily the ones with the most engineers. They're the ones who've thought clearly about what they're building and can translate that into effective agent direction.
If you're a founder or product team looking to move faster with AI coding agents, there are a few concrete ways we can help. We set up the agentic development environments and workflows (Claude Code, Cursor, and custom pipelines) so your team can direct agents effectively from day one rather than burning a month figuring out the tooling. We also build the layer underneath: the APIs, backend services, and integrations that an agent is generating code against need to be well-structured for any of this to work. And where your team has the domain knowledge but not the coding background, we work alongside you to bridge that gap and ship the actual product.
The research says expertise persists. We think that's the right framing for what AI-assisted development actually is: not a replacement for knowing your domain, but a massive multiplier on it. If you want to put that multiplier to work, get in touch.