{
  "video": {
    "id": "ZJjRgvjhq3M",
    "title": "Build Your Own Finance Agent Inside ChatGPT - Stefano Fabbri & Oliver Steffmann, OpenAI",
    "duration": 6632,
    "upload_date": null,
    "channel": "AI Engineer",
    "source": "AI Engineer",
    "event_format": ""
  },
  "analysis": {
    "video_id": "ZJjRgvjhq3M",
    "title": "Build Your Own Finance Agent Inside ChatGPT - Stefano Fabbri & Oliver Steffmann, OpenAI",
    "one_liner": "OpenAI's pitch for finance teams is to build inside the AI surfaces your firm has already approved instead of shipping a standalone app: package your skills, MCP data connections, auth and custom UI as Codex/ChatGPT Work plugins, and enterprise controls can get tools to end users \"in weeks instead of month[s]\".",
    "summary": "Two OpenAI applied AI engineers ran a two-hour workshop on a wealth-management use case. It had two parts. First, ChatGPT for Financial Services: a ChatGPT Work plugin with built-in financial data (SEC filings, delayed market data, news, private company profiles, earnings call transcripts) whose \"data ontology\" was reworked for LLM consumption, plus expert-built skills for banking and research tasks. Second, using Codex to build your own skills, plugins (skills + MCP + tools + UI) and the new plugin extensions that customize the Codex desktop UI. Their argument is that plugins are the superset, and that the open-source Codex harness, which their models are post-trained on, lets builders reach knowledge workers without the usual enterprise approval slog. They also covered cost (pairing smaller models with good skills), compaction, sandbox policies, evals and goal mode.",
    "key_points": [
      "ChatGPT for Financial Services is enterprise-only and currently lives only in the ChatGPT Work web interface. Its financial data cannot be reached from the Codex CLI or a pipeline, and plugin extensions (UI customization) work only on the desktop for now.",
      "Built-in vendors include Pitchbook and Crunchbase, among others. The differentiator is not raw data but pre-processing: each dataset was onboarded with its provider and exposed as \"a really customized tool call for that data set\", with context on what matters in, say, a 10-K. They said explicitly that custom embeddings are not the secret.",
      "The banking and equity research plugins (a library of skills plus MCPs) were built and validated with top investment banks, asset managers and equity researchers. They are meant to give you \"80%\" as a baseline, with your own 20% of customization on top.",
      "Cost strategy: Astra is for genuinely new research or a new codebase. For repeatable tasks, well-built skills let cheaper models (6.1 Soul, Luna) reach top-model performance. In multi-agent setups the orchestrator should be the smartest model, while Terra or Luna subagents get \"basically 95%\" of the result.",
      "A plugin is skills + MCP server + tool configuration + UI components (the app UI standard), zipped into one package. Plugin extensions add custom icons, settings and navigation in the Codex desktop app, popping up context-specific UI such as client-meeting prep.",
      "Dom rebuilt a standalone API-based financial deep-research app as a Codex plugin. The original used a planner agent that splits the question into subquestions, a subagent per subquestion, and a critique agent that spawns new subagents to resolve conflicting findings. As a plugin it uses Codex's dynamic subagents and \"in distribution tool calling\". The point: users had already consolidated their work in Codex and ChatGPT Work.",
      "The Codex harness/app runner is open source and OpenAI's models are post-trained on it. Its compaction mode (a secondary model that compresses context) let 5.5 spend about two hours reproducing a quant paper's codebase (stochastic Brownian motion simulation, model calibration), which they said would be impossible without compaction.",
      "Enterprise admins can push a managed config.toml to set sandbox limits on which URLs, folders and plugins each team can use. A \"guardian\" mechanism automates the critical versus non-critical approval decisions to balance security against convenience."
    ],
    "takeaways": [
      "Before writing a standalone app with an API key, check whether the workflow could ship as a Codex/ChatGPT Work plugin. You inherit login, access control and enterprise approval that way.",
      "Start from vendor or industry skills as a baseline and add your firm's last 20%. Put historical examples (decks, spreadsheets) inside the plugin as reference templates, and fix recurring format errors in the skill itself.",
      "Don't let skills sprawl: define a few core-capability skills per team, prefer plugins when you need controlled deployment, and retest skills whenever a model changes.",
      "Save cost on high-volume work by building strong skills for cheaper models and using a smart orchestrator with cheap subagents, but confirm the quality with evals.",
      "For evals, start simple before reaching for LLM-as-a-judge, and break 1–5 scores into yes/no questions. For hill-climbing, give Codex goal mode a numeric target, and keep auditable decision logs in regulated workflows."
    ],
    "topics": [
      "agents",
      "financial services",
      "plugins",
      "skills",
      "mcp",
      "evals",
      "multi-agent",
      "enterprise security"
    ],
    "tools": [
      "ChatGPT",
      "ChatGPT Work",
      "ChatGPT for Financial Services",
      "Codex",
      "Codex CLI",
      "Codex SDK",
      "Codex app runner",
      "Sites",
      "Plugin Creator",
      "Astra",
      "6.1 Soul",
      "Luna",
      "Terra",
      "5.5",
      "5.6",
      "Responses API",
      "Decision API",
      "goal mode",
      "compaction mode",
      "Daybreak",
      "MCP",
      "Pitchbook",
      "Crunchbase",
      "Salesforce",
      "Figma",
      "Adobe",
      "Jupyter"
    ],
    "quotes": [
      {
        "text": "the orchestrator needs to be like probably like the smartest of the agent",
        "at": "43:35",
        "url": "https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=2615s"
      },
      {
        "text": "advised a customer recently that I think you have too many skills",
        "at": "67:18",
        "url": "https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=4038s"
      },
      {
        "text": "start simple Don't don't start with the LLM as a judge.",
        "at": "100:45",
        "url": "https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=6045s"
      },
      {
        "text": "don't ask to rank one to five to score one to five always ask yes or no it's much better",
        "at": "101:09",
        "url": "https://www.youtube.com/watch?v=ZJjRgvjhq3M&t=6069s"
      }
    ],
    "words": 20514
  },
  "summary_url": "/#ZJjRgvjhq3M",
  "transcript": {
    "html": "/transcripts/ZJjRgvjhq3M.html",
    "txt": "/transcripts/ZJjRgvjhq3M.txt",
    "vtt": "/transcripts/ZJjRgvjhq3M.vtt"
  }
}