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

## Executive summary

The discussion addresses the critical shift in software development from a coding bottleneck to a verification bottleneck, driven by AI agents. Speakers outline the concept of 'software factories,' distinguishing between 'light' (human-guided) and 'dark' (autonomous) models. Key findings emphasize that while AI agents can generate code at unprecedented speeds (e.g., Anthropic scaling to 30,000 merged changes daily), maintaining quality requires adopting multi-layered, zero-trust verification across all development loops (CI, agentic, and supply chain).

## Key takeaways

- Verification is the new bottleneck: The primary challenge is no longer writing code, but verifying the output of agents. The goal is to enable shipping code faster *with confidence* (3:51).
- Adopt Multi-Layered Verification: A zero-trust approach requires verification in every loop. This includes a deterministic layer (catching bugs/SQL injection), an agentic review layer (catching logic/intent issues), and a software supply chain security layer (CVEs, malware) (11:38).
- Software Factories Spectrum: A 'light' factory involves human judgment at every step, while a 'dark' factory is fully autonomous. Most teams are currently somewhere on this spectrum (2:49).
- Accountability and Labeling: While Anthropic is evolving past labeling agent-written code, speakers recommend labeling agent-written PRs in the short term (6-9 months) to maintain auditability and accountability (14:41).

## Technical details

- Anthropic Scaling Metrics: Anthropic scaled its merged code changes from approximately 2,500 to 30,000 per day, utilizing Claude for first-pass review, which includes end-to-end testing, linting, and tool running to build confidence (5:38).
- CI/CD Failure Rates: According to CircleCI, while CI workflows saw a 59% throughput jump, three in ten attempts to merge code into production are still failing (1:57).
- State of Code Survey Findings: The survey found that 84% of companies report AI features shipping faster, yet only 31% of code is reaching production without review, highlighting a significant risk in regulated environments (8:29).
- Agent Control Mechanisms: For enterprise adoption, focus must be placed on control, permissions, and sandboxes. Sonar suggests using 'guide rails' to prevent agents from operating outside defined boundaries (13:19).

## Practical implications

- Implement robust, multi-layered CI/CD pipelines that go beyond simple unit tests, incorporating deterministic, agentic, and supply chain checks.
- Focus development efforts on generating verification evidence (e.g., end-to-end tests, property tests, visual verification) to build confidence in agent-generated code.
- Establish clear governance, permissions, and sandboxing mechanisms ('guide rails') before allowing agents to operate autonomously in production environments.
- Treat the process of shipping code as a controlled assembly line, ensuring that the speed of writing code does not outpace the rigor of verification.

## Topics

AI Engineering, CI/CD, Software Architecture, DevOps, Code Quality, SonarSource, Anthropic, Claude Code, State of Code Survey

Source: https://www.youtube.com/watch?v=AnsXWH2_EBU