How Anthropic Merges 30,000 Code Changes a Day - Manish Kapur, Sonar & Addy Osmani, Anthropic

Highlights

• Anthropic currently scales its operations by merging 30,000 code changes every single day.

• CircleCI analyzes 28 million workflows and finds that throughput increases by 59 percent while 30 percent of merge attempts fail.

• Seventy-six percent of developers generating artificial intelligence code report that production happens at a speed making manual verification impossible.

• Software factories operate on a spectrum ranging from lights-on environments requiring human oversight to dark autonomous factories.

• Giving agents methods to independently verify correctness through tests and browser runs can multiply output quality by two to three times.

Timeline

1. 00:00 - 03:50

   Context and Scaling Software Factories

   Developers increasingly rely on coding agents to open pull requests, but industry data reveals a persistent gap between high code generation throughput and production merge failures. Software factories manage loops at scale, dividing workflows into lights-on models where humans retain oversight and dark models operating via robotics and automation.

   • Coding agents frequently open pull requests across development teams.
   • CircleCI data shows a 59 percent increase in throughput alongside a 30 percent failure rate for production merge attempts.
   • Software factories range from lights-on environments with human judgment to dark factories running autonomously.
2. 03:50 - 08:21

   The Verification Bottleneck

   Code generation speed outpaces manual review capabilities, forcing engineering leaders to prioritize verification and confidence. Anthropic scales its output dramatically by utilizing Claude for initial pull request reviews, running browser tasks, and enforcing automated correction loops to maintain code quality.

   • Eighty-five percent of organizations utilize coding agents for development tasks.
   • Three-quarters of developers state that artificial intelligence generates code too quickly for manual verification.
   • Anthropic scales performance by merging 30,000 code changes daily through automated first-pass reviews.
3. 08:21 - 14:42

   Industry Trends and Multi-Layered Verification

   Survey data indicates that pull requests grow larger and more complex while manual review capacity lags behind. Robust verification strategies rely on multiple layers, catching deterministic security bugs in the first layer and logic issues in the second layer while maintaining strict sandbox controls.

   • Eighty-four percent of surveyed developers report that artificial intelligence delivers features and fixes to production faster.
   • Only five percent of companies currently operate software factories autonomously.
   • Zero-trust, multi-layered verification combines deterministic checks, agentic reviews, and supply chain security.
4. 14:42 - 18:45

   Leadership Strategy and Future Outlook

   Engineering leaders must define verification standards and guide agents within secure sandboxes. Although writing code is largely solved, maintaining speed requires automated verification loops and clear accountability before merging changes into production.

   • Teams should label agent-written code in pull requests to preserve accountability and auditability.
   • Junior developers benefit from mastering diverse testing methodologies including end-to-end and property tests.
   • Strong review frameworks separate teams that ship fast from teams that ship blind.