Agent Nodes: Instructions, Model and Output Format

Narrow instructions and a schema

An agent node wraps a model call with instructions, a chosen model and reasoning settings, optional tools, and an output format (free text or a structured JSON schema). Give each agent one narrow responsibility (classify, extract, answer from documents, write a reply) rather than one giant prompt. When another node consumes the result, use structured output so fields like category and urgent can be routed on reliably. In code, the same idea appears as an Agent with an output_type.

A structured output type in the Agents SDK, run

I ran this with Python 3, the openai-agents package 0.22.3 and pydantic 2.13.5, without any API key: it only builds definitions and schemas, no model is called. A pydantic model becomes the agent's output schema: three required fields, with category limited to billing, technical or other. A JSON reply parses into a typed object.

import json
from typing import Literal
from pydantic import BaseModel
from agents import Agent, AgentOutputSchema

class Triage(BaseModel):
    category: Literal["billing", "technical", "other"]
    urgent: bool
    summary: str

agent = Agent(name="Classifier", instructions="Classify the ticket.", output_type=Triage)
schema = AgentOutputSchema(Triage).json_schema()
print("required fields:", schema["required"])
print("category enum  :", schema["properties"]["category"]["enum"])
print("parsed example :", Triage.model_validate_json('{"category": "billing", "urgent": false, "summary": "double charge"}'))

Output:

required fields: ['category', 'urgent', 'summary']
category enum  : ['billing', 'technical', 'other']
parsed example : category='billing' urgent=False summary='double charge'

Use smaller models for simple nodes

Classification or extraction nodes often work well on a smaller, cheaper model; save large models for the steps that need reasoning.

Quick check: Why give a classifier agent a structured output schema?

• To make the reply longer
• So later nodes can branch on typed fields reliably
• Because text output is impossible
• To skip testing