A look at Anthropic's first developer conference, where CPO Mike Krieger said 70%+ of Anthropic's code pull requests are now written by their AI model Claude
www.zdnet.com/article/firs... Kylie Robison / @kylierobison.com : if you were not at anthropic's very first developer conference - dont fret - i wrote about literally everything that happened on and off livestream www.wired.com/story/anthro... Karl Bode / @karlbode.com : salesman say thing he sell do good big things [embedded post] X: Kylie Robison / @kyliebytes : @DarioAmodei interpretability! AI personality! policy! product vision! biomedical research! AGENTS! read all the anthropic takes and enjoy my photo dump that didn't make it on wired dot com https://www.wired.com/... [image] Kylie Robison / @kyliebytes : here's what @DarioAmodei said about President Trump's megabill that would ban state-level AI regulation for 10 years https://www.wired.com/... [image] Kyle Russell / @kylebrussell : “About 80% of people at Anthropic that are technical use Claude Code every day.” [image] Alex Albert / @alexalbert__ : The demand for Claude 4 already in just 24 hours is absolutely insane. My favorite part is hearing from all the startups on the cutting edge who've now tried Opus and suddenly their product just... works.
Context & Ripple Effects
Anthropic made its Claude Code command-line agent generally available alongside Claude 4 updates, including thinking summaries, in the release immediately preceding the conference. The company is now presenting internal use as evidence that the tool is becoming part of its own engineering workflow.
That internal adoption sits alongside Anthropic's longer-standing effort to differentiate its product development through safety-focused decision-making, documented in earlier reporting on its safety-oriented culture. The pull-request claim gives developers a concrete, if company-reported, measure of how far AI-assisted coding has moved inside the lab.
First-order effects
- Anthropic can use the reported 70%+ Claude-authored pull-request share as a high-profile validation signal for Claude Code just as it seeks developer adoption.
- Anthropic engineers' work shifts further toward specifying, reviewing, debugging, and integrating model-produced changes rather than drafting every change manually.
Second-order effects
- Competing coding-model vendors face added pressure to demonstrate reliable use in production engineering workflows, not only benchmark performance.
- Early startup demand for Claude 4 may strengthen the feedback loop between external developer usage and Anthropic's product iteration, especially around agentic coding tools.
Third-order effects
- If AI-generated pull requests become routine, software teams may reorganize around review, testing, and governance capacity; code generation itself becomes less of the bottleneck.
- The durable competitive advantage may move from a standalone model to an integrated coding workflow that earns engineers' trust at scale, though Anthropic's internal usage is not by itself proof of customer-wide outcomes.
The trend: This is one data point in the shift from AI coding assistants as optional autocomplete to agentic systems embedded in the software delivery process.