Anthropic demo shows 90% self-written code

Anthropic has been documenting an internal software workflow in which its coding agent, Claude Code, is used to write large portions of the product itself. The company said in a January 29, 2026 product post that internal adoption of Claude Code was associated with a 67% increase in pull requests merged per engineer per day, and that across teams, 70% to 90% of code was being written with Claude Code assistance. That makes the recent social-media attention around a demo of “self-written” code less of an isolated claim than a public retelling of figures Anthropic has already published. Anthropic has also separately described Claude Code as an “agentic coding system” that can read a codebase, make changes across files, run tests and deliver committed code. ### Where does the 90% figure come from? Anthropic’s own published number is not “fully autonomous” in the literal sense of no human role. (claude.com) In the January 29 post, the company said “70–90% of code is now being written with Claude Code assistance,” which frames the metric as assisted code generation inside a human-managed engineering workflow. A separate Anthropic research post from December 2, 2025 said the company surveyed 132 engineers and researchers, conducted 53 interviews and studied internal Claude Code usage data. (claude.com) That post said employees reported using Claude in 59% of their daily work and seeing average productivity gains of 50%, while also noting concerns about maintaining technical depth and supervising outputs. ### What exactly is Claude Code doing inside the workflow? (claude.com) Anthropic’s product materials describe Claude Code as more than an autocomplete tool. The product page says it can understand a codebase, edit files, run commands and handle development tasks across a project. Anthropic’s engineering posts describe a broader agent pattern in which models call tools through code, use loops and conditionals, and work across testing frameworks, package managers, git operations and deployment pipelines. (anthropic.com) That matters because the company’s published throughput figure is tied to a workflow, not just to faster text generation. Anthropic said contribution metrics are calculated by matching Claude Code session activity with GitHub commits and pull requests, and that the data is intended as one input alongside other engineering metrics. ### Does this mean Anthropic engineers are no longer coding? Mike Krieger, Anthropic’s chief product officer, said in a February 2026 appearance at Cisco’s AI Summit that for most products at Anthropic, Claude was “effectively 100%” of the code-writing process, according to reports citing the event. (anthropic.com) Those reports said engineers were shifting toward directing, reviewing and approving machine-generated code rather than writing every line by hand. (claude.com) Anthropic’s own research is more measured. The December 2025 post said engineers were becoming more productive and more “full-stack,” but it also said some worried they could become less able to supervise Claude’s outputs or lose deeper technical competence. ### What is still unclear about the demo claim? The public Anthropic sources support the broad direction of the claim, but they do not fully define the denominator behind “90% self-written code.” Anthropic’s published wording is “with Claude Code assistance,” and its metrics post says the company counts only code where it has “high confidence” in Claude Code’s involvement. (officechai.com) The company did not, in the materials reviewed, publish a line-by-line methodology for separating human-written, AI-assisted and AI-generated code. (anthropic.com) The same limitation applies to the throughput number. Anthropic said pull requests are an incomplete measure of developer velocity, even as it uses merged PRs per engineer per day as a practical proxy for shipping work. ### What should readers watch next? Anthropic already exposes some of this story in public-facing materials. The company’s recorded webinar pages and product documentation continue to show how Claude Code is used in live workflows, while its engineering blog has expanded into longer-horizon agent design, including harnesses for autonomous software tasks. (claude.com) The next useful data point will likely be a more detailed methodology release, another internal research update, or fresh product analytics from Anthropic on how much code is assisted, reviewed and merged by teams using Claude Code at scale. (claude.com) (anthropic.com)