{
  "video": {
    "id": "Bdrs3uAX0_M",
    "title": "AI Coding Agents Are Breaking Big Codebases — Dan Adler, Sourcegraph",
    "duration": 743,
    "upload_date": null,
    "channel": "AI Engineer",
    "source": "AI Engineer",
    "event_format": ""
  },
  "analysis": {
    "video_id": "Bdrs3uAX0_M",
    "title": "AI Coding Agents Are Breaking Big Codebases — Dan Adler, Sourcegraph",
    "one_liner": "Dan Adler argues that coding agents are producing a flood of code that is making huge, decades-old enterprise codebases decay, and that agents need infrastructure built for that scale: a code graph, search, and compiler-accurate analysis. His line for it is \"Visibility is infrastructure.\"",
    "summary": "Adler (Sourcegraph CEO) says most software engineers work in large, long-lived codebases with thousands of repositories that \"keep the world running\". AI coding agents are now adding code to these codebases faster than teams can review or maintain it. The volume causes duplicated code, inconsistent standards from different agents, more brittle cross-service dependencies and new vulnerabilities. He says agent builders (OpenAI, Anthropic, Cursor) are not building the infrastructure needed to see and understand a 50,000-repo codebase. He presents Sourcegraph's code graph and its new agentic batch changes product (in beta) as the answer, with an example from an early user at Mercari.",
    "key_points": [
      "72% of software industry employment is in companies with more than 500 people, so most engineers work in large, long-lived codebases with thousands of repositories, not at startups.",
      "Agents produce a \"tidal wave\" of code (\"title wave\" in the transcript). The volume itself is the problem: \"The call is coming from inside the house\". The tools that speed developers up also create the conditions that make these codebases fail.",
      "Symptoms he lists: different agents applying different coding standards; duplicated code because the agent doesn't know a library already exists; more brittle cross-service dependencies; small deviations from standards that cause hidden issues; and new vulnerabilities uncovered by agents every day.",
      "Codebases with millions of lines and tens of thousands of repositories are too big to fit in a context window, or even to clone and grep in real time.",
      "LLMs \"just love to search\" and will grep their way to understanding a codebase, so repo architecture written into AGENTS.md files is \"barely moving the needle\". Understanding is a context problem: an agent can't grep code it cannot see.",
      "Customer anecdotes: a tech leader at a top-10 car manufacturer overheard a developer say \"I don't know what this code does. AI wrote it for me\" (about vehicle autopilot code). A tech leader at a top-10 US bank faced an npm supply-chain vulnerability across 90,000 repositories.",
      "Sourcegraph's agentic batch changes product launched in beta that day. From a single prompt it rolls out changes iteratively across hundreds or thousands of repositories. It self-heals by responding to CI status and PR comments, uses coding agents where judgment is needed and deterministic scripts where it isn't, and tracks coverage so you can be sure 100% of affected places are patched.",
      "Mercari example: a user fixed a GitHub code injection issue (environment variables had to be set correctly) in two known repos, then had the tool explore the rest of the codebase. It found 80 other potential vulnerabilities and patched them in one go with a deterministic script."
    ],
    "takeaways": [
      "Don't rely on AGENTS.md architecture notes to give agents codebase understanding. Give them search and code-graph tooling across all repositories, because agents can't grep what they can't see.",
      "For changes that span many repositories (for example supply-chain vulnerability patches), use deterministic scripts where no judgment is needed and agents only where it is, and track coverage until every affected location is fixed.",
      "Expect AI-generated code to cause duplication, drift in standards and brittle cross-service dependencies, and plan for active oversight of legacy codebases instead of only adding review agents as \"sandbags\".",
      "Know what is actually running in production: which repos, forks, copies, and external repos pulled in as libraries or APIs."
    ],
    "topics": [
      "coding agents",
      "large codebases",
      "code search",
      "code graph",
      "context",
      "batch changes",
      "security vulnerabilities",
      "enterprise"
    ],
    "tools": [
      "Sourcegraph",
      "Sourcegraph agentic batch changes",
      "Claude Code",
      "OpenAI",
      "Anthropic",
      "Cursor",
      "AGENTS.md",
      "npm",
      "GitHub",
      "Mercari"
    ],
    "quotes": [
      {
        "text": "Sure, Claude Code can uh make this change, but I have 90,000 repositories to make it in.",
        "at": "06:45",
        "url": "https://www.youtube.com/watch?v=Bdrs3uAX0_M&t=405s"
      },
      {
        "text": "I don't know what this code does. AI wrote it for me.",
        "at": "04:46",
        "url": "https://www.youtube.com/watch?v=Bdrs3uAX0_M&t=286s"
      },
      {
        "text": "But understanding is a context problem. You can't gp what you literally cannot see.",
        "at": "07:52",
        "url": "https://www.youtube.com/watch?v=Bdrs3uAX0_M&t=472s"
      },
      {
        "text": "Visibility is infrastructure.",
        "at": "09:18",
        "url": "https://www.youtube.com/watch?v=Bdrs3uAX0_M&t=558s"
      }
    ],
    "words": 2328
  },
  "summary_url": "/#Bdrs3uAX0_M",
  "transcript": {
    "html": "/transcripts/Bdrs3uAX0_M.html",
    "txt": "/transcripts/Bdrs3uAX0_M.txt",
    "vtt": "/transcripts/Bdrs3uAX0_M.vtt"
  }
}