Claude Code data shows expertise, not coding job title, drives agent success

Anthropic’s latest Claude Code research shifts the coding-agent debate from demos to usage data. In a June 16 tweet, Anthropic wrote: “Our latest economic research introduces a framework for tracking Claude Code as it scales. Who is using Claude Code, and what are they using it for? How is the value of tasks changing? And how much does domain expertise shape whether a session succeeds?” The source tweet is here.

The linked report studies about 400,000 interactive Claude Code sessions from roughly 235,000 people between October 2025 and April 2026. Anthropic says it uses privacy-preserving transcript classifiers and telemetry rather than reading individual sessions. The analysis focuses on command-line, Claude.ai, and Claude Code desktop usage, excluding third-party IDE integrations, SDK usage, and headless CLI prompts.

The headline finding is that people still make most planning decisions while Claude makes most execution decisions. Anthropic estimates users make about 70% of planning decisions but only 20% of execution decisions. A typical session has around four turns, and each user prompt triggers about 10 Claude actions on average. Expert-rated sessions go further: roughly 12 actions and 3,200 words of output per prompt, compared with about five actions and 600 words for novice sessions.

The economic signal is also concrete. Anthropic estimates the average task value rose 27% from October to April, using a comparison to freelance job postings as a relative measure. The work mix changed as well: fixing broken code fell from 33% to 19% of sessions, while operating software rose from 14% to 21%, and writing plus data analysis roughly doubled from 10% to 20%.

Anthropic’s account usually publishes Claude product updates, safety research, and economic analysis. This tweet is material because it gives a quantitative answer to whether coding agents only help software engineers. In code-producing sessions, every one of the ten largest occupations in Anthropic’s dataset landed within seven percentage points of software/math users on success. The next thing to watch is whether those gaps shrink further as agents improve, or whether domain expertise remains the main advantage for getting useful work from them.