{
  "video": {
    "id": "x4e5O9zN0TE",
    "title": "Long-Horizon Agents Need Experiments, Not Just Prompts — Erina Karati",
    "duration": 1287,
    "upload_date": null,
    "channel": "AI Engineer",
    "source": "AI Engineer",
    "event_format": ""
  },
  "analysis": {
    "video_id": "x4e5O9zN0TE",
    "title": "Long-Horizon Agents Need Experiments, Not Just Prompts — Erina Karati",
    "one_liner": "To improve stateful agents over long horizons, stop hand-tuning prompts. Freeze the harness, scenarios and metrics, then let an autoresearch loop make small edits to a narrow policy surface such as memory writing, retrieval, trust and source attribution. Score whole runs on a balanced scorecard, and keep a change only if the scorecard improves and the guardrails hold.",
    "summary": "The speaker built Project Paradox at Supercell's AI Innovation Lab. It is a modular multi-agent framework for video game companions. Each agent has its own RAG memory, an emotion vector, trust scores toward other agents and the player, and importance-scored memories. It worked for short-term gameplay, but over longer horizons social consistency degraded: rumors lost their source, turned into facts, or were forgotten during planning. Inspired by Karpathy's autoresearch, the team wrapped the village in a meta-level experiment loop. The loop runs controlled scenarios, collects traces, scores behavior against ground truth on a multi-metric scorecard, proposes small changes to the agent protocol, and keeps only the ones that improve behavior across the whole society. The speaker argues that the same pattern applies to any agent that keeps state over time.",
    "key_points": [
      "Project Paradox architecture has four stateful parts. (1) Each agent has its own memory namespace backed by RAG, so memories don't bleed between agents. (2) Emotion is a small vector (joy, sadness, fear, anger, disgust) updated after events. (3) Belief/trust scores toward other agents and the player act as a trust matrix; after each interaction the LLM decides whether a score goes up, down or stays the same. (4) Each memory gets an LLM-assigned importance score, and memories above a threshold go into a separate cache so they can be retrieved better later.",
      "Failure over long horizons: in the 'mango sale' rumor demo, a rumor passed from agent to agent. After other events had happened, the agent the player asked failed to give the expected context. Typical failures: the agent keeps the rough topic but loses its source, a rumor is stated as fact, or the agent knows a fact but doesn't use it when planning.",
      "Autoresearch is a meta-system outside the village, not another villager. It reads full run traces, compares them with the scenario's ground truth, scores the behavior, proposes a constrained change to the agent protocol or cognitive policy, and reruns. The key shift is evaluating an entire run instead of one answer.",
      "Controlled scenario suites make fuzzy social behavior measurable. Examples: public fact diffusion ('the bakery will close tomorrow'), rumor uncertainty (does 'might leave' become 'is leaving'?), and replanning (a group learns its planned route is blocked).",
      "Use a balanced scorecard, not a single 'agent quality' metric: diffusion reach after N steps, provenance/source retention, uncertainty preservation and false-certainty rate for rumors, action consistency and time-to-replan for planning, and containment for privacy. Optimizing only diffusion teaches agents to overshare; optimizing only recall produces noisy or stale memories.",
      "Keep the editable surface small. The harness, scenarios and metrics are frozen. Only these are exposed: memory-writing policy, retrieval policy, communication prompt, belief/trust rules, source attribution and replanning triggers. That is the difference between the LLM 'writing random patches' and 'searching within a controlled policy space'.",
      "Example policy changes: if source attribution disappears, keep the source in memory writes and summaries. If rumors harden into facts, store confidence, mark firsthand vs secondhand, and require hedging when retelling. If public facts stay local, classify useful public facts differently and have agents share them proactively.",
      "After running an autoresearch loop, the mango-rumor agent answered in context after a long period. The speaker was careful about claims: without repeated loop results they would not say the system 'generally improved', only that this is the right kind of surface to expose."
    ],
    "takeaways": [
      "Follow the recipe: freeze the harness, define scenarios, log traces, score behavior, expose only a small policy surface, search over changes, and keep only the changes that survive measurement.",
      "Build controlled scenario suites with ground truth for long-running agent behavior. Don't judge from free-roaming demos or 'vibes'.",
      "Score with a balanced, multi-metric scorecard that includes guardrails such as privacy containment, so the optimizer can't game one number.",
      "Treat rollback as mandatory. Run the loop as a ratchet: try a change, score it, and keep it only if the scorecard improves and the guardrails hold, because a gain in one metric (for example diffusion) can damage another (for example private-information leakage).",
      "Memory alone is not enough. Store provenance (who said it; firsthand, secondhand, verified or uncertain) and keep raw episodic memories separate from what the agent currently believes."
    ],
    "topics": [
      "agents",
      "multi-agent",
      "evals",
      "autoresearch",
      "agent memory",
      "rag",
      "game ai",
      "long-horizon"
    ],
    "tools": [
      "Project Paradox",
      "Supercell",
      "Microsoft",
      "Andrej Karpathy's autoresearch",
      "RAG"
    ],
    "quotes": [
      {
        "text": "Instead you want a balanced scorecard.",
        "at": "13:22",
        "url": "https://www.youtube.com/watch?v=x4e5O9zN0TE&t=802s"
      },
      {
        "text": "This is the difference between the LM writing random patches versus the LM actually searching within a controlled policy space.",
        "at": "15:29",
        "url": "https://www.youtube.com/watch?v=x4e5O9zN0TE&t=929s"
      },
      {
        "text": "And the biggest lesson for me perhaps was that memory is not enough here.",
        "at": "17:05",
        "url": "https://www.youtube.com/watch?v=x4e5O9zN0TE&t=1025s"
      },
      {
        "text": "Long horizon agents need experiments and not just prompts.",
        "at": "20:55",
        "url": "https://www.youtube.com/watch?v=x4e5O9zN0TE&t=1255s"
      }
    ],
    "words": 3157
  },
  "summary_url": "/#x4e5O9zN0TE",
  "transcript": {
    "html": "/transcripts/x4e5O9zN0TE.html",
    "txt": "/transcripts/x4e5O9zN0TE.txt",
    "vtt": "/transcripts/x4e5O9zN0TE.vtt"
  }
}