{
  "video": {
    "id": "8G_1-3IO4ZQ",
    "title": "WTF Is the Context Layer? The Missing Infrastructure for Production Agents — Prukalpa Sankar",
    "duration": 1254,
    "upload_date": null,
    "channel": "AI Engineer",
    "source": "AI Engineer",
    "event_format": "In person"
  },
  "analysis": {
    "video_id": "8G_1-3IO4ZQ",
    "title": "WTF Is the Context Layer? The Missing Infrastructure for Production Agents — Prukalpa Sankar",
    "one_liner": "Atlan's founder argues that model intelligence has 1,000x'd while business context has barely moved, and walks through two failed-then-refined internal eras — bootstrapped per-task agents, then a shared \"company brain\" of 300 skills — to define the context layer as versioned, dependency-managed, self-learning infrastructure: \"the GitHub for context.\"",
    "summary": "Prukalpa Sankar frames real-world agent performance as intelligence × context, noting that intelligence has 1,000x'd in a decade (2x in the last 6 months) while situated business knowledge sits untouched in dashboards, Slack threads and an analyst's head — which is why only 1 in 5 AI use cases reach production and 56% of CEOs report zero financial benefit. She narrates Atlan's own two eras: 18 months of bootstrapping single-purpose agents (Hermione, MoneyPenny) that hit context-engineering cost, island agents, untraceable errors and context sprawl across Relevance → Google ADK → Glean → Claude Code → Claude/Codex; then a shared context layer where domain experts author skills into one company brain serving general-purpose agents. That second approach produced ~300 skills and 40 agents in the marketing team in 6 months, but exposed that context must be managed like code — dependency breakage, drift, unclear ownership, hardcoded .env secrets. Her close: today we build agents by hardcoding context, the scale of that problem is underhyped, and context is not just king but IP.",
    "key_points": [
      "Performance = intelligence + context; only 10% of human job performance variance is explained by IQ, and models are getting exponentially smarter without getting exponentially more useful — 1 in 5 AI use cases reach production, 56% of CEOs report zero financial benefit from AI.",
      "Intelligence has 1,000x'd in a decade and 2x'd in the last 6 months; context — 'the situated knowledge of your business' — has barely moved beyond some data reaching the cloud.",
      "The 'Maya the analyst' example decomposes a simple question ('why is my drive-thru time up this week?') into knowledge (metric definitions, week cut-off, timezone), expertise (diagnostic playbooks — check Q3 seasonality, check last quarter's product launch), and norms (persona scoping — who's asking).",
      "Era one at Atlan (~18 months ago): a jobs-to-be-done map of the customer-experience role, a 'scaling factor' per task, then bootstrapped single-purpose agents — Hermione (health intelligence lead), MoneyPenny (financial risk analyst). Building an agent took 5 minutes; giving it enough business context to be accurate 'took forever'.",
      "Era one's failure modes: agents on their own island (marketing changed positioning, the website SDR agent kept pitching the old version), no traceability when an agent got it wrong (model? agent? context?), and context sprawl from per-agent memory systems — compounded by tooling churn across Relevance, Google ADK, Glean, Claude Code and now ~50/50 Claude and Codex, with context trapped in each.",
      "Era two: a shared context layer / 'company brain' — the best SEO person writes the SEO skill, the best competitive-intel person writes the competitive-intel skill, into a common repo with retrieval mechanisms feeding general-purpose agents. Atlan's marketing team produced ~300 skills and 40 agents in 6 months; the brain needed a data graph, a skills library, semantics/metrics (what is ARR, what is a qualified lead) and org structure/entities.",
      "Era two's failure modes prove context must be managed like code: the competitive-intelligence skill feeds category positioning which feeds sales battle cards, so each skill's learning breaks something downstream; skills drift and go stale; nobody owns skill quality; governance was 'a nightmare' with secrets hardcoded in .env files and people downloading public skill repos.",
      "Proposed shape of a context layer: continuously mine context from business systems, feed one company brain, manage skills through a context development lifecycle with versioning, approvers/maintainers/contributors and security posture, expose it via MCP, SQL, vector retrieval and hybrid assembly, and close the loop with a harness that reads agent traces and hands reconstructions back to maintainers as approve/reject."
    ],
    "takeaways": [
      "Stop treating context as a per-agent prompt artefact — a general-purpose agent fed from one shared brain beats N single-purpose agents with N private memory systems, especially since your agent framework will change several times a year and take its trapped context with it.",
      "Apply code practices to context: versioning, dependency mapping (which downstream skills break when this one learns), named approvers/maintainers/contributors, quality ownership and security posture — and get secrets out of .env files and public skill repos.",
      "Build the compounding learning loop from traces: deploy a specialized harness that reverse-constructs improvements from agent traces and surfaces them to a human maintainer as approve/reject rather than expecting agents to self-improve silently.",
      "To bootstrap from a sprawl of 60 systems, don't start by writing docs — connect Salesforce/HubSpot to the warehouse to the application layer and reverse-construct how they link, since context is lost at every hop; that reconstruction yields an accurate first version of the company brain.",
      "Treat the encoded context as differentiating IP, not overhead: when you and your competitor have the same models, knowledge, playbooks and norms are what distinguish your customer support agent from theirs."
    ],
    "topics": [
      "agents",
      "context-engineering",
      "skills",
      "knowledge-graphs",
      "data-governance",
      "multi-agent-systems",
      "memory",
      "enterprise-ai"
    ],
    "tools": [
      "Atlan",
      "Claude Code",
      "Codex",
      "Google ADK",
      "Glean",
      "Relevance",
      "Co-work",
      "Qualified",
      "Artisan",
      "Salesforce",
      "HubSpot",
      "Slack",
      "MCP",
      "GitLab",
      "Zoom",
      "Discord",
      "Affirm",
      "MasterCard",
      "General Motors"
    ],
    "quotes": [
      {
        "text": "There is no doubt that the models are getting exponentially smarter by the day. 2 years ago they couldn't pass the bar. Today, if they were to take the bar, it was they're the top 1% of test scorers. On the other hand, they're not exponentially more useful by any benchmark.",
        "at": "02:09",
        "url": "https://www.youtube.com/watch?v=8G_1-3IO4ZQ&t=129s"
      },
      {
        "text": "we got to the point by middle of last year where building an agent was really easy. Took like 5 minutes. Uh but giving it the business context that it took to actually get it to be accurate took forever.",
        "at": "08:48",
        "url": "https://www.youtube.com/watch?v=8G_1-3IO4ZQ&t=528s"
      },
      {
        "text": "The question I like to ask is what does the GitHub for context look like?",
        "at": "15:39",
        "url": "https://www.youtube.com/watch?v=8G_1-3IO4ZQ&t=939s"
      },
      {
        "text": "Today, we're largely building agents by hardcoding context. The scale of this problem I truly believe is underhyped, because with scale this can become really unsustainable.",
        "at": "18:58",
        "url": "https://www.youtube.com/watch?v=8G_1-3IO4ZQ&t=1138s"
      }
    ],
    "words": 3938
  },
  "summary_url": "/#8G_1-3IO4ZQ",
  "transcript": {
    "html": "/transcripts/8G_1-3IO4ZQ.html",
    "txt": "/transcripts/8G_1-3IO4ZQ.txt",
    "vtt": "/transcripts/8G_1-3IO4ZQ.vtt"
  }
}