OpenAI's Agent Builder

Summary of technological concepts & product features (Agent Builder / Agent Kit)

• Agent Builder is presented as a drag-and-drop workflow system for building AI agents, conceptually similar to tools like Make.com or other visual node builders.
• Templates are a core starting point, providing example agent designs with predefined node structures.

Workflow structure (nodes + conditional routing)

Workflows are built from nodes connected like a graph (conceptually similar to LangGraph):

• Nodes perform steps such as:
  • guardrails
  • classification
  • tool calls
  • agent actions
• Edges are often conditional, functioning like an if/else router that determines the next node based on structured outputs.
• Nodes can incorporate tools to extend capabilities and influence behavior.

Example template: Structured “Data Q&A” style agent

A demonstrated template includes:

• Guard rails to control unsafe prompt behavior.
• A classification step that routes requests into categories such as:
  • Q&A (internal Q&A)
  • Fact-finding (external search/tools)
  • Other (ask the user for more details or route elsewhere)
• Structured JSON outputs from the classifier, including fields like an “operating procedure” enum.

Tools highlighted in the templates / system

Web search agent

• Uses a model referenced as GPT-5 / “GPT5” to search the web.
• Returns text with configurable verbosity (e.g., medium).

External fact-finding routing

• Can include:
  • Web search
  • Code interpreter tool

File search tool

• Loads documents into a vector store (described as a “mini RAG”).
• Allows configuring the number of returned results (per the described settings).

Guard rails

Multiple guardrail types are mentioned:

• Jailbreak prevention guardrail
  • Not perfect, but described as easier than handcrafted approaches.
• Moderation guardrail
  • Blocks harmful content.
• Optional hallucination grounding behavior using a vector store.

MCP support

• Ability to call external MCP servers, including:
  • “ones by OpenAI”
  • other developers’ MCP servers.

---

Tutorial / hands-on build: “Westworld travel agent” workflow

1. Create new workflow and drag in agents/steps.
2. Add:
   • Start node
   • Guard rails (uses the jailbreak guardrail in the demo)
   • Classifier agent with JSON schema output, including an enum:
     • Westworld
     • other
     • non-travel
   • Classifier prompt instructs the model to classify the user’s topic and output structured data.
3. Add an if/else router based on the classifier’s JSON field (e.g., user topic).
4. Configure three downstream behaviors:
   • If Westworld:
     • Route to a Westworld-specific travel agent
     • Use a vector store built from a Markdown “master guide” document (pricing + itinerary + details)
     • Instruct the agent to:
       • answer from the document without false info
       • be persuasive and friendly
     • Demo behavior includes generating a 3-day itinerary and pulling pricing from the document.
   • If other (e.g., Disneyland):
     • The agent politely declines and redirects to “Delos customer support.”
   • Non-travel / unsupported topic:
     • Redirects similarly (the enum includes this category, even if not deeply demonstrated).
5. The demo shows:
   • Correct routing and retrieval for Westworld questions
   • Proper rejection for Disneyland queries

---

Deployment & integration/exports

• The workflow can be published as a named agent (e.g., “Westworld travel agent”).
• Deployment options mentioned:
  • ChatKit (noted as interesting; a separate video is promised)
  • Using an Agents SDK to export/generate code for running elsewhere
• Additional flexibility notes:
  • The creator suggests the SDK may allow changing models
  • It may limit usage to OpenAI-specific features such as:
    • file search
    • guard rails (per subtitle commentary)

---

Strengths and stated weaknesses (review/analysis)

Strengths

• Quick onboarding via templates, visual workflow building, and built-in components (guardrails, file search/RAG, routing, tools, MCP).
• Clear demonstration of:
  • structured classification
  • conditional routing
  • RAG grounded answers from uploaded documents
• “Export it out” via SDK is positioned as helpful for adoption.

Downsides / concerns

• Possible lock-in to the OpenAI ecosystem, especially through ChatKit / Agent SDK constraints.
• The speaker advocates for an open-source, framework-agnostic builder that could target multiple agent ecosystems (e.g., LangGraph, ADK, PydanticAI, etc.).
• Notes that MCP support could reduce wiring complexity by integrating external capabilities.

---

Main speaker / sources (from the subtitles)

• Main speaker: The video’s narrator/creator (first-person “I” throughout; no named identity provided in the subtitles).
• Primary product/source referenced: OpenAI “Agent Builder” / “Agent Kit” / Agents SDK / ChatKit, plus related examples involving:
  • guardrails
  • file search vector stores
  • MCP integration