An agent persona is the set of instructions that defines an agent’s role and guides how it handles tasks and responds. Often called the system prompt, it can specify expertise, communication style, domain focus, and operating constraints. A writing assistant and a coding assistant may use different personas even if both can answer questions. section 1.3.1
Think of the persona as the agent’s operating brief, not its entire skill set. Tools let an agent take actions; reasoning and planning help it choose what to do. The persona shapes those choices and constrains how the result is communicated. These parts interact throughout a task rather than running in a fixed sequence. section 1.3
Write a persona as a clear contract:
- Name one focused role and state what the agent does and does not do. Include an explicit way to say “I don’t know” when information is missing. section 11.1.1
- Specify the output format when another program or agent will consume the answer. A JSON response matched to a typed schema avoids brittle string parsing. section 11.1.1
- Supply relevant changing facts, such as the date or organization, through runtime instructions rather than hardcoding them. Too much injected context can add tokens and confuse the agent. section 11.1.1
A focused support persona
A support assistant can be instructed to answer concisely, prefer retrieved context, admit when context is missing, and never invent facts or make policy exceptions. Its response can follow a typed format with an answer, citations, and a confidence value. Each instruction has a job: set behavior, establish a boundary, or make the result easier for software to handle. section 11.1.1
Listing 11.1 01_core_persona_tips.py
from pydantic import BaseModel, Field
from agents import Agent, Runner
from datetime import date
class Answer(BaseModel): #1
"""Machine-readable answer format."""
answer: str = Field(..., description="Concise, user-facing answer")
citations: list[str] = Field(default_factory=list)
confidence: float = Field(..., ge=0, le=1)
def core_instructions(ctx, agent) -> str: #2
today = date.today().isoformat() #3
return (
f"You are SupportMentor, a helpful domain assistant. Today is {today}.\n" #4
"Always:\n"
"1) Answer concisely; 2) Prefer retrieved context; 3) If context is missing, say you don't know.\n"
"Output must be valid JSON matching the Answer schema.\n"
"Never: make policy exceptions or invent facts."
)
core_agent = Agent(
name="SupportMentor",
instructions=core_instructions,
output_type=Answer
)
input = "What's our refund window for accessories?"
result = Runner.run_sync(core_agent, input)
print(result.final_output)
The listing shows how to combine runtime instructions with a Pydantic response model and an OpenAI Agents SDK agent. section 11.1.1
