{
  "video": {
    "id": "cKhpeEBnT1o",
    "title": "Your Coding Agent Is 6 Months Out of Date — Jakub Hojsan, Exa",
    "duration": 743,
    "upload_date": null,
    "channel": "AI Engineer",
    "source": "AI Engineer",
    "event_format": ""
  },
  "analysis": {
    "video_id": "cKhpeEBnT1o",
    "title": "Your Coding Agent Is 6 Months Out of Date — Jakub Hojsan, Exa",
    "one_liner": "Coding and code-review agents work about 6 months behind because of model knowledge cutoffs. Giving them web search only helps if you also give them rules on when to use it, such as checking the upstream changelog whenever a diff bumps a dependency.",
    "summary": "Jakub Hojsan, a forward deployed engineer at Exa, argues that LLMs can't properly write or review code that depends on changes made after their knowledge cutoff, which typically trails release by about 6 months. His example is a PR that removes a parameter. Without search, a model reads it as a cleanup; with search, it finds the dependency bump and the breaking changes in the upstream changelog. He pitches Exa's semantic search, with token-efficient highlights and full trace transparency, as the fix, and argues that bolting a search tool onto an agent is not enough without rules for when to use it. He also introduces Exa Agent, which adds data partners such as SimilarWeb, Particle and Crunchbase.",
    "key_points": [
      "Models typically lag about 6 months between knowledge cutoff and release. Repo changes made in that window fall into the model's 'blind spot': it can't build with them and can't review them.",
      "Demo: a PR removing a parameter ('inertia check') from a vector store repo, made after GPT-5.5's cutoff. Without web search the LLM calls the diff a consistent cleanup. With search it finds the upstream changelog's breaking changes, sees the change is a refactor for a dependency bump, and explains the migration.",
      "Exa doesn't return the whole GitHub page. An interpretation step boils the page down to the snippet the LLM needs, so the model gets about 500 characters instead of 100,000.",
      "Highlights are pulled at runtime by computation, not by an LLM, so Hojsan says they add no extra latency. The same page gives different highlights for different queries; his demo used his own bio page, asking about hobbies and then for his phone number.",
      "Adding a search tool isn't enough. Claude Code has native web search, yet it will still say Sonnet 4.6 or 4.7 doesn't exist. Agents need explicit rules about when to search, then they need to actually run the search.",
      "Exa's four claimed advantages over native OpenAI or Anthropic web search: transparency (the full trace of queries, sources and highlights, which can be fed into telemetry, versus a black box that can take up to 10 seconds and returns sources but no content), token-efficient highlights, lower cost on large workloads, and no lock-in to one model provider (one fixed API across GLM, Anthropic and OpenAI models).",
      "Cursor, Cognition, Warp and CodeRabbit use Exa for web search. The new Exa Agent adds partner data: SimilarWeb (web analytics), Particle (podcast intelligence) and Crunchbase (private markets). It is used by financial firms and hedge funds.",
      "Q&A: retrieval is multi-stage (query embedding, keyword filtering plus semantic search, reranking, then dropping irrelevant results) over a highly curated index of 'tens of billions' of documents. Structured output schemas can be generated from natural language, with up to 10 fields on 'deep' and he believes up to 100 on agents."
    ],
    "takeaways": [
      "Give code-review agents explicit rules. For example: when a diff bumps a dependency or version, check the upstream source or changelog and base the review on what you find.",
      "Don't assume a model knows about anything released after its cutoff. Expect wrong answers about new library versions and new models unless the agent is told to search.",
      "Feed the model short, query-specific excerpts (about 500 characters) instead of whole pages, so you don't waste context and don't need an LLM summarisation step.",
      "Prefer a search setup where you can see the exact queries, sources and highlights, and send that trace into your telemetry so you can debug bad answers.",
      "A third-party search API keeps search the same across model providers, so switching models doesn't change your retrieval layer."
    ],
    "topics": [
      "coding agents",
      "code review",
      "web search",
      "knowledge cutoff",
      "context efficiency",
      "retrieval",
      "observability",
      "agent instructions"
    ],
    "tools": [
      "Exa",
      "Exa Agent",
      "Exa MCP",
      "Claude Code",
      "GPT-5.5",
      "Claude Sonnet 4.6/4.7",
      "OpenAI",
      "Anthropic",
      "GLM",
      "Stack Overflow",
      "GitHub",
      "Google",
      "Cursor",
      "Cognition",
      "Warp",
      "CodeRabbit",
      "SimilarWeb",
      "Particle",
      "Crunchbase",
      "Nvidia"
    ],
    "quotes": [
      {
        "text": "coding agents don't really need these 10 blue links that you see when you search Google.",
        "at": "00:29",
        "url": "https://www.youtube.com/watch?v=cKhpeEBnT1o&t=29s"
      },
      {
        "text": "So you're not feeding the model 100,000 characters anymore. You're feeding it quite literally only 500 characters to answer the question.",
        "at": "03:09",
        "url": "https://www.youtube.com/watch?v=cKhpeEBnT1o&t=189s"
      },
      {
        "text": "Now adding a search tool to your agent is not enough.",
        "at": "03:20",
        "url": "https://www.youtube.com/watch?v=cKhpeEBnT1o&t=200s"
      },
      {
        "text": "So you actually have to provide the model with a set of rules.",
        "at": "03:52",
        "url": "https://www.youtube.com/watch?v=cKhpeEBnT1o&t=232s"
      }
    ],
    "words": 2466
  },
  "summary_url": "/#cKhpeEBnT1o",
  "transcript": {
    "html": "/transcripts/cKhpeEBnT1o.html",
    "txt": "/transcripts/cKhpeEBnT1o.txt",
    "vtt": "/transcripts/cKhpeEBnT1o.vtt"
  }
}