{
  "video": {
    "id": "eXA2tjRZIbY",
    "title": "The 6 Pillars of an Agentic Harness for Production — Varun Krovvidi, Resolve AI",
    "duration": 1265,
    "upload_date": null,
    "channel": "AI Engineer",
    "source": "AI Engineer",
    "event_format": ""
  },
  "analysis": {
    "video_id": "eXA2tjRZIbY",
    "title": "The 6 Pillars of an Agentic Harness for Production — Varun Krovvidi, Resolve AI",
    "one_liner": "Pointing the biggest model at production data demos well but breaks as you scale. Resolve AI says a production agent harness needs six pillars: model orchestration, context engineering, causal reasoning, governed actions, a learning system and evals.",
    "summary": "Varun Krovvidi argues that AI has transformed coding because code is self-documenting, modular and single-domain. Running and fixing production software is different: it is multi-domain, multi-tool and multiplayer, and he says it takes about 70% of engineering time. As setups scale across use cases and teams, the same failures keep appearing: anchoring bias and the model treadmill, context that makes models over- or under-explore, models that give coherent answers instead of causal ones, unsafe actions, and no learning between investigations. Resolve AI's harness answers these with six pillars, which he demos with an alert that Grafana posts to Slack. The agent finds a root cause, lists the theories it ruled out, holds its position when the user pushes back, and pulls teammates into a virtual war room.",
    "key_points": [
      "He describes three kinds of production work using a healthcare analogy: on-call and maintenance (Tylenol/Advil), all-hands incidents (surgery), and ongoing health work such as infra, cost and platform engineering (daily vitamins). Resolve maps these to on-call agents, incident agents and ambient background agents, all running on one shared agent architecture.",
      "Model orchestration has two layers: keeping up with the model treadmill, and matching the best model to each task within a workflow. His examples are Gemini for image reasoning, OpenAI for deterministic steps and Claude for open-ended investigations, and he says this needs constant re-evaluation.",
      "Context engineering is about the precise amount of context needed, not the implementation (graph RAG, knowledge graphs). Too much context makes models over-explore and invent hypotheses; too little makes them under-explore. Telemetry is 'literally infinite'. He describes a mix: graph-RAG-style retrieval to start an investigation, then precise tool calls (log queries, metrics, dashboards, code queries) so it doesn't burn through tokens.",
      "Causal reasoning: a root cause must rest on a causal chain of evidence. If the chain can't be established, the system reports low confidence and points the user in a different direction. It is also designed not to give in when a user pushes back.",
      "Governed actions: least-privilege guardrails that depend on the organisation and industry (read vs write, and under what conditions writes are allowed). He presents an AI deleting a database as 'a feature, not a bug' of models choosing the cleanest fix, so guardrails are required.",
      "Learning system: every interaction is a learning opportunity, including positive reinforcement, negative reinforcement and steering. The context of earlier investigations should carry over so the 100th investigation doesn't start from scratch.",
      "Evals, which he calls the start and end point of any agent architecture, have five levels: positive reinforcement, negative reinforcement, tracing the agent's path, scoring it against how the best engineer would investigate, and confidence calibration. The architecture is re-run against them whenever a model, use case or architecture changes.",
      "Demo: Grafana fires an alert into Slack. Resolve treats the alert as the prompt and spins up investigator agents that reason across code, infrastructure, knowledge bases and observability. It traces a log skill's high failure rate to stale integrations that were still live. It lists ruled-out theories, including a GCP outage at the same time and traffic spikes. When pushed toward a GCP outage or deployment errors as the cause, it sticks to its causal chain."
    ],
    "takeaways": [
      "Don't tie one model to a whole workflow. Build an orchestration layer that routes each sub-task (images, SQL, logs, deterministic steps, open-ended reasoning) to the best model and re-evaluates as new models ship.",
      "Design context as a precise budget: broad retrieval to start an investigation, then narrow, precise tool calls, rather than filling the window with telemetry.",
      "Require a causal chain of evidence before stating a root cause. Otherwise report low confidence, show the ruled-out theories, and don't let the model give in to user pressure.",
      "Give agents least-privilege access, with explicit read/write rules that fit your organisation and industry.",
      "Treat user interactions (approvals, rejections, steering) as eval signals. Score agent paths against expert engineers and against confidence calibration, and re-run evals on every model, use-case or architecture change."
    ],
    "topics": [
      "agents",
      "agent harness",
      "sre",
      "incident response",
      "context engineering",
      "evals",
      "model orchestration",
      "guardrails"
    ],
    "tools": [
      "Resolve AI",
      "Gemini",
      "OpenAI",
      "Claude",
      "Grafana",
      "Slack",
      "Microsoft Teams",
      "GCP",
      "Microsoft",
      "graph RAG",
      "knowledge graph"
    ],
    "quotes": [
      {
        "text": "most people dub it as a last mile but that is the single longest mile that you have to run with AI",
        "at": "05:18",
        "url": "https://www.youtube.com/watch?v=eXA2tjRZIbY&t=318s"
      },
      {
        "text": "So, models are specifically designed to give you coherent answers, but not causality.",
        "at": "10:02",
        "url": "https://www.youtube.com/watch?v=eXA2tjRZIbY&t=602s"
      },
      {
        "text": "AI might have determined that the cleanest possible fix is just deletion of that particular code snippet. You can't fault it for it.",
        "at": "10:47",
        "url": "https://www.youtube.com/watch?v=eXA2tjRZIbY&t=647s"
      },
      {
        "text": "the era of bottomless AI is over like the unlimited AI is no longer there",
        "at": "04:14",
        "url": "https://www.youtube.com/watch?v=eXA2tjRZIbY&t=254s"
      }
    ],
    "words": 4456
  },
  "summary_url": "/#eXA2tjRZIbY",
  "transcript": {
    "html": "/transcripts/eXA2tjRZIbY.html",
    "txt": "/transcripts/eXA2tjRZIbY.txt",
    "vtt": "/transcripts/eXA2tjRZIbY.vtt"
  }
}