OpenAI Agent Builder Workflows
Design agent workflows on a visual canvas and in code: nodes, state, routing, loops, handoffs, approvals, guardrails, evals, versioning and cost, with runnable Agents SDK and Python snippets.
What you'll learn
- Explain where a visual agent workflow builder fits next to the Agents SDK, chat UI embedding and evaluation tools.
- Design a workflow from start, agent, tool, logic and end nodes with clear state and structured outputs.
- Route requests with if/else branches, bounded loops, handoffs and human approval steps.
- Add guardrails, least-privilege tools and defences against prompt injection.
- Test workflows with previews, traces and evaluation datasets before publishing a version.
- Estimate cost per run and operate published workflows with versioning and monitoring.
Syllabus
What Agent Builder Is and When to Use It
- Agent Builder in the OpenAI Platform
- Visual Builder or Code?
- The Workflow Mental Model: Nodes, Edges and State
Core Building Blocks
- The Start Node: Inputs and State Variables
- Agent Nodes: Instructions, Model and Output Format
- Tools: File Search, MCP and Functions
- End Nodes and Workflow Outputs
Logic: Branches, Loops and Approvals
Data, State and Templating
Multi-Agent Patterns
Guardrails and Safety
- Guardrail Nodes
- Least Privilege for Tools and MCP Servers
- Prompt Injection Through Files, Web and Tools
Testing, Evals, Versions and Deployment
- Preview Runs and Traces
- Evaluating Workflows With Datasets
- Publishing and Versioning Workflows
- Deploying: Embedded Chat or Exported Code