Anthropic's Claude Code Accelerates Developer Workflows

The headline number is 90% of Anthropic's code being machine-written, but the more durable signal for practitioners is what Anthropic had to build around that fact: a load-bearing automated code-review layer, an ownership model that keeps a human accountable for every merged change, and a feedback loop tight enough to auto-generate its own bug-fix pull requests. Teams adopting agentic coding at any scale will hit the same review bottleneck before they hit any generation-quality ceiling.

O'Reilly reports that Cat Wu, who leads product for Claude Code and Cowork at Anthropic, said 90% of Anthropic's codebase is now written by Claude Code. The account comes from a fireside chat Wu held with Addy Osmani at O'Reilly's AI Codecon, written up by Tim O'Reilly on O'Reilly Radar (April 30, 2026). Per the writeup, Anthropic's internal "dog-fooding" Slack channel for Claude Code receives a new message every 5 to 10 minutes, and engineers ship new internal builds of the tool multiple times a day. O'Reilly recounts one specific episode: Wu introduced a small interaction bug between prompts and auto-suggestions, and before she got to fixing it, a teammate's scheduled Claude Code task - set to scan the feedback channel for anything unaddressed after 24 hours - had already found the issue and opened a pull request to fix it.

Wu told O'Reilly that Anthropic engineers are now producing roughly 200% more code than a year earlier, and that the resulting bottleneck is review, not generation. Her team's response was to run the most comprehensive automated review configuration it could, tuning for high bug recall rather than a lighter touch, on the reasoning that a review system only earns trust as a load-bearing part of the pipeline if it is thorough. That internal practice lines up with Anthropic's separately shipped Code Review feature for Claude Code: the company has reported that substantive review comments on pull requests rose from 16% to 54% after deploying it, with engineers disagreeing with fewer than 1% of the tool's findings, per Anthropic's own product materials. O'Reilly's account adds that the review agents trace changes across files, catching bugs unrelated to the immediate diff - such as a cache-invalidation bug surfaced during an unrelated encryption refactor.

Three separate layers are doing distinct work here, and teams building agentic coding pipelines should instrument each one rather than treating "AI writes the code" as a single capability: the generation layer (the agent producing edits), the orchestration layer (scheduled or triggered agents that monitor channels and open PRs on their own), and the verification layer (automated review tuned for high recall, plus a human owner accountable end-to-end for each merged change, including post-deploy bugs). Anthropic's account suggests the ownership model matters as much as the tooling: PR authors, not reviewers, are on the hook for what ships, which is what makes it acceptable to lean on heavy automated review in the first place.

Formal provenance capture for agent-authored changes (signed metadata, reproducible prompts), broader adoption of test-oracle patterns that check intent rather than surface output, and operational guardrails - rate limits, merge gating, security scanning - as more teams let scheduled agents open PRs unattended the way Anthropic's internal tooling already does.