{
  "video": {
    "id": "AnsXWH2_EBU",
    "title": "How Anthropic Merges 30,000 Code Changes a Day - Manish Kapur, Sonar & Addy Osmani, Anthropic",
    "duration": 1133,
    "upload_date": null,
    "channel": "AI Engineer",
    "source": "AI Engineer",
    "event_format": ""
  },
  "analysis": {
    "video_id": "AnsXWH2_EBU",
    "title": "How Anthropic Merges 30,000 Code Changes a Day - Manish Kapur, Sonar & Addy Osmani, Anthropic",
    "one_liner": "Anthropic went from about 2,500 merged changes a day in May to about 30,000 by having Claude do a first-pass review that actually exercises the change like a user would, with hooks that correct the PR on every edit and humans checking the agent's verification evidence rather than every line.",
    "summary": "Manish Kapur of Sonar and Addy Osmani of Anthropic argue that writing code with AI is now largely solved, so the bottleneck has moved to verification and 'code confidence'. Osmani lays out a spectrum from 'lights-on' software factories, where humans apply judgment and taste at each step, to 'dark' factories, where agents scope, build and check while humans only handle intent and rare exceptions. He describes how Anthropic reviews code: Claude does a first pass with end-to-end and browser checks, hooks run on every edit, team taste is codified in Claude.md or agents.md, and humans check the agent's evidence. Kapur presents Sonar's case for zero-trust, multi-layered verification inside every loop: a deterministic layer, an agentic review layer and a supply-chain security layer.",
    "key_points": [
      "CircleCI analysed 28 million CI workflows this year: throughput rose about 59%, yet three in 10 attempts to merge code into production fail.",
      "Anthropic merges about 30,000 code changes a day, up from about 2,500 in May. The Claude.ai refresh, which a case study says roughly tripled its performance, took about 3,000 pull requests in a short time from engineers, product designers and other functions.",
      "Anthropic's review workflow: Claude does a first-pass review and tries the experience the way a real user would (end-to-end tasks, checking it in a browser, running tools and linting). Hooks run on every edit and study the diffs so the PR gets corrected quickly. Anthropic also uses its own code review product.",
      "Team best practice and taste should be codified in a shared Claude.md or agents.md that the agent consults during review. Humans are then responsible for checking the agent's evidence, such as browser-test videos of it clicking around, or mobile and desktop screenshots.",
      "Osmani says giving agents a way to verify that work is correct and complete can easily improve the quality of outcomes two to three times. He says the next gains will come from nailing verification, because model quality (e.g. Opus 5.5) is already pretty good.",
      "Sonar's State of Code survey (more than 1,000 developers): 84% say AI gets features and fixes to production faster, PRs are getting larger and more complex, and 31% more code reaches production without review. 5% of companies use software factories today, 34% are expected to, and 39% in financial services. 28% have approved agents with shared controls.",
      "Sonar's multi-layered, zero-trust verification has three layers: deterministic analysis for bugs and security issues such as SQL injection and buffer overflows; agentic review for logic and intent issues; and supply-chain security for CVEs, malware and malicious packages. It applies in the agentic loop, the CI loop and code maintenance.",
      "Enterprises that let teams spin up hundreds or thousands of agents worry about agent identity, telemetry, control, permissions and sandbox escape. Sonar frames its answer as 'guide rails' rather than guardrails."
    ],
    "takeaways": [
      "Give your agents a way to verify their own work, through end-to-end tests, property tests, visual verification, browser runs, screenshots and videos, and review that evidence instead of reading every line.",
      "Write your team's taste and best practices into Claude.md or agents.md so automated code review can apply them.",
      "Put independent, automated, layered verification (deterministic, agentic and supply-chain) inside every loop: the agentic loop, CI and maintenance.",
      "Label agent-written PRs, at least for the next 6 to 9 months, for auditability and to trigger extra checks. Keep a named human owner who is accountable for each change.",
      "Leaders should work out the gap between their current testing and letting agents sign off on merges. Juniors should build experience with the different kinds of tests (end-to-end, property and others)."
    ],
    "topics": [
      "code review",
      "verification",
      "software factories",
      "agents",
      "testing",
      "ai governance",
      "security",
      "ci/cd"
    ],
    "tools": [
      "Claude Code",
      "Claude",
      "Claude.ai",
      "Claude Opus 5.5",
      "Anthropic",
      "Sonar",
      "CircleCI",
      "Claude.md",
      "agents.md",
      "Google"
    ],
    "quotes": [
      {
        "text": "A software factory is harnessing loops at scale.",
        "at": "03:03",
        "url": "https://www.youtube.com/watch?v=AnsXWH2_EBU&t=183s"
      },
      {
        "text": "you need to give your agents a way to verify that the work is correct.",
        "at": "05:14",
        "url": "https://www.youtube.com/watch?v=AnsXWH2_EBU&t=314s"
      },
      {
        "text": "basically, you want to get to a place where the agent's giving you as much verification evidence as possible.",
        "at": "08:15",
        "url": "https://www.youtube.com/watch?v=AnsXWH2_EBU&t=495s"
      },
      {
        "text": "Fast to write is not fast to ship.",
        "at": "17:58",
        "url": "https://www.youtube.com/watch?v=AnsXWH2_EBU&t=1078s"
      }
    ],
    "words": 4266
  },
  "summary_url": "/#AnsXWH2_EBU",
  "transcript": {
    "html": "/transcripts/AnsXWH2_EBU.html",
    "txt": "/transcripts/AnsXWH2_EBU.txt",
    "vtt": "/transcripts/AnsXWH2_EBU.vtt"
  }
}