{
  "video": {
    "id": "KcVkq5L-0f0",
    "title": "Your agents lack context: Here's how to fix \"You're absolutely right!\" — Brandon Waselnuk, Unblocked",
    "duration": 849,
    "upload_date": null,
    "channel": "AI Engineer",
    "source": "AI Engineer"
  },
  "analysis": {
    "video_id": "KcVkq5L-0f0",
    "title": "Your agents lack context: Here's how to fix \"You're absolutely right!\" — Brandon Waselnuk, Unblocked",
    "one_liner": "Brandon Waselnuk of Unblocked argues the bottleneck is no longer model intelligence but context, and shows that feeding an agent a real \"context engine\" cut the same task from ~21M tokens to 10.8M and saved about two hours of wall clock.",
    "summary": "Waselnuk frames every engineer as having spent years being a \"context engine\" — built from asking questions, getting PRs rejected and being on call — and points out that a fresh agent session is intelligent but has none of it. He argues the two common fixes are local maxima: the \"curated context trap\" of hand-written markdown repos that rot and need an omnipotent curator, and the \"MCP plateau\", where the agent either never calls the tool or stops at the first plausible hit due to satisfaction-of-search bias. Instead he specifies a context engine with six properties — unified system context, targeted retrieval, conflict resolution, personalized relevance, token optimization and permission enforcement — and shows A/B numbers from running the same prompt on the same model with and without it. He closes by giving away three open-source resources: a GitHub social-graph tool, a repo rules agent, and a workshop workbook on building a relational context engine beyond RAG.",
    "key_points": [
      "Same prompt, same model, with vs. without context: roughly 21 million tokens without versus 10.8 million with, about 2 hours of wall-clock time saved, and better answer quality — headline claim of ~50% fewer tokens and faster triage, because the wasted search tokens spent rediscovering the codebase at the start of every session disappear.",
      "The cost of bad context compounds along the adoption curve: tab-complete (cheap, human vetoes it instantly) → agents without a human in the loop → doom loops of correct-correct-correct that burn search tokens and rework time → an AI code-review tax → fully background agents that must be able to query for answers themselves.",
      "Failure mode 1, the \"curated context trap\": markdown files the agent greps do work at first, but then you must distribute them (a GitHub repo), the repo rots like every other doc, and someone must be the omnipotent taste-maker curating it for the whole org.",
      "Failure mode 2, the \"MCP plateau\": depending on how you write the server and tool descriptions the agent may never call it, and if it does, satisfaction-of-search bias means it takes the first thing that looks right — finding an architecture record and never seeing last night's Slack thread saying do A instead of B.",
      "Six properties of a context engine: unified system context across all sources (so it surfaces your unknown unknowns), targeted retrieval fast enough to unfurl a link (deep research when you want it, speed when you need it), conflict resolution between an old architecture diagram and last night's CTO Slack message, personalized relevance (who you are, where your commits are, who reviews them), token-optimized responses for machine-to-machine calls, and permission enforcement via OAuth/SSO so secret project A doesn't leak.",
      "Open-source tool 1 — a \"social network\" tool that runs deterministically over your GitHub to map who commits where and who reviews whose work, producing a distilled experts graph; adding an OpenAI or Anthropic API key optionally labels your teams for you.",
      "Open-source tool 2 — a repo rules agent that discovers every place your team has written rules files, reports severities and duplicate or conflicting rules, and produces a grep-able index so you can dedupe and improve retrieval.",
      "Workshop workbook (six stacked PRs) on going beyond RAG to a relational context engine: RAG alone cannot answer \"what are the open PRs I worked on in the last week with authentication?\" — that needs queries, so the technique is a schema-less lookup where the agent discovers a schema and writes deterministic queries against it.",
      "Use cases go past code generation: customer success staff resolving tickets as they arrive and salespeople closing deals earlier in the quarter by querying the Unblocked context engine in the field."
    ],
    "takeaways": [
      "Shift context left the same way you shift defects left — fix bad context at the start of a session, because every wrong assumption compounds into doom loops, wasted search tokens and rework further down the agentic curve.",
      "Stop treating a curated markdown repo or a bolted-on MCP server as the answer; assume the agent will stop at the first plausible source, and design so it must reconcile conflicting sources rather than settle for one.",
      "Give agents personalized, permission-scoped context (who the requester is, where they commit, who reviews them) and token-optimized responses for machine-to-machine calls, keeping richer prose for the humans who query the same engine from Slack.",
      "Pair RAG with a relational query path: let the agent discover a schema and write deterministic queries so it can answer relational questions RAG cannot.",
      "Run the free tools on your own org — the GitHub social-graph tool to see who really owns what, and the repo rules agent to find duplicate and conflicting rules files — and take the readiness.unblocked.com quiz to see which level of the curve you're on."
    ],
    "topics": [
      "context-engineering",
      "agents",
      "rag",
      "mcp",
      "tokenomics",
      "code-review",
      "developer-tools",
      "enterprise-codebases"
    ],
    "tools": [
      "Unblocked",
      "MCP",
      "GitHub",
      "Slack",
      "OpenAI API",
      "Anthropic API",
      "Fable",
      "readiness.unblocked.com",
      "LinkedIn",
      "Workday",
      "General Motors"
    ],
    "quotes": [
      {
        "text": "AI-generated code should feel like it was written by someone who's been on your team for years.",
        "at": "01:13",
        "url": "https://www.youtube.com/watch?v=KcVkq5L-0f0&t=73s"
      },
      {
        "text": "The problem here is access to information is not understanding.",
        "at": "05:42",
        "url": "https://www.youtube.com/watch?v=KcVkq5L-0f0&t=342s"
      },
      {
        "text": "what your agent can't see is everything below the waterline. It can 100% get code that compiles, but that code that compiles is taking down prod and you have a P0 at 1:00 in the morning.",
        "at": "05:54",
        "url": "https://www.youtube.com/watch?v=KcVkq5L-0f0&t=354s"
      },
      {
        "text": "The gap is not intelligence any longer. It's context.",
        "at": "13:11",
        "url": "https://www.youtube.com/watch?v=KcVkq5L-0f0&t=791s"
      }
    ],
    "words": 3209
  },
  "summary_url": "/#KcVkq5L-0f0",
  "transcript": {
    "html": "/transcripts/KcVkq5L-0f0.html",
    "txt": "/transcripts/KcVkq5L-0f0.txt",
    "vtt": "/transcripts/KcVkq5L-0f0.vtt"
  }
}