Does an AI agent that can act on its own change how people delegate work, compared with a chat interface? Anthropic's Economic Index report of 28 April 2025, "AI’s impact on software development", asks this for coding, the use that the first Anthropic Economic Index report found to be far more common on Claude than the share of computer-related jobs in the US workforce would predict. The post compares conversations on Claude.ai, the chat product, with sessions on Claude Code, then described as Anthropic's new specialist coding agent that can carry out chains of tasks with digital tools. The post argues that coding may be a leading indicator for how other occupations change as agents spread, while noting that lessons from software may not carry over directly ("Looking ahead").
What the post does
Data
The analysis covers 500,000 coding-related interactions, split evenly between the two products before reweighting (section "How we analyzed conversations"; footnote 1). The footnote gives the details:
- all conversations are from 6 to 13 April 2025;
- Claude.ai conversations are from Free and Pro accounts only, with a Claude-based filter applied to select coding conversations;
- Claude Code sessions are only those running on Anthropic's own API, not on third-party cloud providers;
- results are renormalized so that the two products are weighted equally where the post combines them.
Method
Conversations were analyzed with Clio (Clio: A system for privacy-preserving insights into real-world AI use), Anthropic's tool for producing aggregated, anonymized descriptions of conversations. Clio assigned each conversation a topic (for example "UI/UX component development"), a programming language, a project type, and an interaction pattern.
The interaction patterns follow the earlier Economic Index framework. Automation means the AI directly performs the task; augmentation means the AI collaborates with the user. Each splits into subtypes, defined in the caption to the post's first figure:
- Directive (automation): complete delegation with minimal interaction.
- Feedback Loop (automation): task completion guided by feedback from the environment, for example the user pasting error messages back to Claude.
- Task Iteration (augmentation): collaborative refinement.
- Learning (augmentation): the user acquiring knowledge or understanding.
- Validation (augmentation): checking and improving the user's work.
What it finds
Automation is more common on the agent
On Claude Code, 79% of conversations were classified as automation and 21% as augmentation; on Claude.ai, 49% were automation (key findings, item 1; "How do developers interact with Claude?"). Within automation, Feedback Loop conversations were 35.8% on Claude Code against 21.3% on Claude.ai, and Directive conversations 43.8% against 27.5%. Every augmentation pattern, including Learning, was substantially lower on Claude Code (same section).
The post qualifies the result in two ways. Feedback Loop interactions still involve a person, even if only to paste errors, so humans are "still very often involved" within automation; the post says it is not certain this will persist as agents need less input (same section). Its limitations add that the line between automation and augmentation "becomes increasingly blurred" with agentic tools and that the framework will probably need extending ("Limitations").
An appendix compares software and non-software conversations on Claude.ai alone. Software work is more automative, driven by an increase in Feedback Loops (given as +18.3%) that offsets a decrease in Directive behavior (-11.2%); the post reads this as coding requiring a lot of human review and iteration compared with non-coding tasks ("Appendix").
What developers build
Across both products (equally weighted), JavaScript and TypeScript together accounted for 31% of queries and HTML and CSS another 28%; Python was 14% and SQL 6% ("What are developers building with Claude?"). The post notes that Python and SQL probably include much data analysis as well as back-end work, and that HTML figures on Claude.ai are likely slightly inflated because Artifacts use HTML (footnote 2). Among tasks, "UI/UX Component Development" accounted for 12% and "Web & Mobile App Development" for 8% of conversations, both among the top five. The post speculates that jobs centered on simple applications and interfaces "might face earlier disruption" if "vibe coding" (describing the desired outcome in natural language and letting the AI handle implementation) becomes mainstream.
Who uses it, a more tentative analysis
The post labels its project-type analysis "more preliminary" than the rest, because it relies on inferences from incomplete context ("Who is using Claude for coding?"). In that section, startup work accounted for 32.9% of Claude Code conversations, described as nearly 20% higher than its share on Claude.ai, and enterprise work for 23.8% of Claude Code conversations, slightly below its 25.9% share on Claude.ai. Students, academics, personal projects and tutorial or learning uses together made up half of interactions across both products. The key-findings list at the top of the post gives the startup and enterprise figures differently, as 33% of Claude Code conversations serving startup work against "only 13%" for enterprise-relevant applications; the post does not explain the difference between the 13% and 23.8% figures. Footnote 3 adds that enterprise use of Claude.ai is likely undercounted, because Team and Enterprise plans were excluded.
Limitations stated by the post
The "Limitations" section lists:
- only Claude.ai (Free and Pro) and Claude Code were analyzed, not Team, Enterprise or API use;
- the automation and augmentation boundary blurs for agentic tools;
- project-type labels were educated guesses; a "Could Not Classify" option was chosen for 5% of Claude.ai and 2% of Claude Code conversations, which were excluded before renormalizing;
- the users are likely early adopters;
- privacy limits confined the data to a specific retention window (the conversations are from 6 to 13 April 2025), which may miss cycles such as sprints and releases;
- developers use other tools, so the view is partial;
- the study covers what is delegated, not how outputs are used, the quality of the resulting code, or productivity.
Research questions
- Does the Feedback Loop share fall as agents improve? The post asks this itself. Repeating the same classification on Claude Code sessions from later model generations would show whether human-in-the-loop automation turns into Directive automation.
- How reliable are Clio's interaction-pattern labels for agentic sessions? The post says the categories blur for agents. Human annotation of a sample of Claude Code sessions, compared with Clio's labels, would measure agreement for each subtype.
- Which figure describes enterprise use of Claude Code? The post gives 13% and 23.8%. Published definitions for the two categories, or a re-run with a single definition, would resolve which share applies.
- Does automation share predict output quality? The post did not study code quality or productivity. Linking interaction pattern to downstream outcomes, such as whether generated code was kept, would show whether more delegation means more useful work.
- Is the web-front-end skew a property of the work or of the users? The sample is likely early adopters. Comparing language and task distributions with an independent survey of professional developers would separate the two.
How it connects
- The post builds on The Anthropic Economic Index (February 2025), which introduced the automation and augmentation framework and the finding of disproportionate computer-related use.
- It uses Clio for all classification, as do Values in the Wild and How people use Claude for support, advice, and companionship.
- It cites the education report Anthropic Education Report: How university students use Claude for the parallel finding that Computer Science students show highly disproportionate use.
References
- Anthropic. (2025, April 28). Anthropic Economic Index: AI’s impact on software development. https://www.anthropic.com/research/impact-software-development