Agentic coding and persistent returns to expertise

Worauf sich diese Seite beruft, wörtlich, abgerufen am 12.10.2026:

• Writing and data analysis roughly doubled, from about 10% to 20% of sessions.bestätigt 10.11.2026
• We classify each session by having a model read its transcriptbestätigt 10.11.2026
• ~400,000 Claude Code sessions from between October 2025 and April 2026bestätigt 10.11.2026
• or where code is incidental to the final productbestätigt 10.11.2026
• and 13% produce analysis or prosebestätigt 10.11.2026
• interactive sessions from ~235,000 people between October 2025 and April 2026bestätigt 10.11.2026
• About 56% of sessions consist of writing (25%), fixing (26%), or testing and orchestrating code (5%).bestätigt 10.11.2026
• Operating software comprises 17%, while 14% of sessions are planning or exploring, and 13% produce analysis or prosebestätigt 10.11.2026
• the share of sessions spent fixing broken code fell from 33% to 19%bestätigt 10.11.2026
• Operating software grew from 14% to 21% of sessions.bestätigt 10.11.2026
• usage shifted toward more end-to-end agentic usebestätigt 10.11.2026
• the estimated value of the average session rose by 27% between October and Aprilbestätigt 10.11.2026
• These price estimates are coarsebestätigt 10.11.2026
• On average, people make about 70% of the planning decisions but only 20% of the execution decisionsbestätigt 10.11.2026
• people decide what to build, and the agent decides how to build itbestätigt 10.11.2026
• In a typical session, there are about four such turns.bestätigt 10.11.2026
• each prompt the user sends sets off a chain of around 10 actions taken by Claude on averagebestätigt 10.11.2026
• A senior engineer asking their first Rust question is a beginner at Rust.bestätigt 10.11.2026
• 15% of the time and at least partial success 77% of the timebestätigt 10.11.2026
• A session rated intermediate or up reaches verified success 28-33% of the timebestätigt 10.11.2026
• a working grasp of the domain captures most of the benefitbestätigt 10.11.2026
• every one of the ten largest occupations in our dataset lands within seven points of software engineers in terms of their successbestätigt 10.11.2026
• may now be able to do technical work they previously could notbestätigt 10.11.2026
• These findings are preliminary.bestätigt 10.11.2026
• via presentations and other prose-based documentsbestätigt 10.11.2026
• writes on average 2,400 words of outputbestätigt 10.11.2026
• Claude takes fewer actions per turn (about eight actions)bestätigt 10.11.2026
• it takes on the highest number of actions (about 16)bestätigt 10.11.2026
• About 2% of sessions average more than 100 actions per promptbestätigt 10.11.2026
• In typical novice sessions, each prompt sets off about five Claude actions and roughly 600 words of outputbestätigt 10.11.2026
• expert sessions set off action chains more than twice as long (12 actions) carrying five times the output (3,200 words)bestätigt 10.11.2026
• A session rated intermediate or up reaches verified success 28-33% of the time and partial success 91-92% of the timebestätigt 10.11.2026
• rises from 4% for novice-rated sessions to 15% for expert-rated onesbestätigt 10.11.2026
• 19% of sessions where the user appears to be a novice end abandoned, against 5-7% for everyone elsebestätigt 10.11.2026
• We were able to infer occupation in about 70% of sessions.bestätigt 10.11.2026
• The fastest-growing non-software occupation groups in our sample are management, sales, and legal occupations.bestätigt 10.11.2026
• People in software-related occupations reach verified success in about 30% of their sessions overallbestätigt 10.11.2026
• those numbers are 34% and 29% respectivelybestätigt 10.11.2026
• reaching at least partial success in code-producing sessions 89% and 88% of the timebestätigt 10.11.2026
• it has neither widened nor narrowed over seven monthsbestätigt 10.11.2026
• Management occupations are highest on verified success, slightly above the software engineering occupations.bestätigt 10.11.2026
• managers may be more likely to communicate when they get what they ask forbestätigt 10.11.2026
• much of it is programmatic, with Claude Code embedded in automated tools and pipelines rather than conversing with a userbestätigt 10.11.2026
• the non-interactive usage this report excludes is a substantial share of activitybestätigt 10.11.2026
• All classifiers in this report use Claude Sonnet 4.6 unless otherwise noted.bestätigt 10.11.2026
• No researcher reads individual transcriptsbestätigt 10.11.2026
• more than 90% of sessions our classifier labeled as creating or modifying code showed code changes in the telemetrybestätigt 10.11.2026
• But classifiers remain challenging to validate at scalebestätigt 10.11.2026
• we cannot measure real-world outcomes, like whether code written in a session is actually used or discarded thereafterbestätigt 10.11.2026
• Experts hit trouble less often overall, so the troubled sessions they do have are likely to be on harder problemsbestätigt 10.11.2026
• novices get stuck on routine problems while experts get stuck on challenging hard problemsbestätigt 10.11.2026
• rose in almost every kind of workbestätigt 10.11.2026
• about 25% on averagebestätigt 10.11.2026
• all grew more valuable by roughly a third or more (about 43%, 34%, and 32% respectively)bestätigt 10.11.2026
• Claude Code users now spend an average of 20 hours per week using the toolbestätigt 10.11.2026
• measures hours in which Claude Code was actively runningbestätigt 10.11.2026
• The picture in this report will be updated as the models, the users, and the division of labor between them change.bestätigt 10.11.2026
• if the returns to expertise begin to decrease over timebestätigt 10.11.2026
• so we use them primarily to compare tasks to one another over timebestätigt 10.11.2026
• catches the edge case it mishandles at month-end closebestätigt 10.11.2026
• Each point is an adjusted ratebestätigt 10.11.2026
• that would suggest that models are starting to supply the essential judgment that users currently bringbestätigt 10.11.2026
• understanding how an existing system works, and planning a change before making itbestätigt 10.11.2026
• deploying, configuring, running pipelines, monitoring systemsbestätigt 10.11.2026
• Building, operating, and fixing-type tasks all grew more valuable by roughly a third or more (about 43%, 34%, and 32% respectively).bestätigt 10.11.2026