Validation

Validate agent responses and retry on failure.

Basic usage

import json

def is_valid_json(response: str) -> bool:
    """Check if response is valid JSON."""
    try:
        json.loads(response)
        return True
    except:
        return False

agent = Agent(
    validator=is_valid_json,
    validation_retries=3,
)

result = await agent.run("Return a JSON object with name and age")
# Guaranteed to be valid JSON (or raises after 3 retries)

How it works

1. Agent generates response
2. Validator function is called with the response
3. If validator returns True, response is returned
4. If validator returns False:
   • Agent retries with feedback
   • Repeats up to validation_retries times
5. If all retries fail, last response is returned

Validator function

The validator must:

• Accept a single str argument (the response)
• Return bool (True = valid, False = invalid)

def validator(response: str) -> bool:
    # Your validation logic
    return True  # or False

Examples

JSON with required fields

def has_required_fields(response: str) -> bool:
    try:
        data = json.loads(response)
        return "name" in data and "email" in data
    except:
        return False

agent = Agent(validator=has_required_fields, validation_retries=2)

Minimum length

def min_length(response: str) -> bool:
    return len(response) >= 100

agent = Agent(validator=min_length, validation_retries=2)

Contains keywords

def contains_summary(response: str) -> bool:
    return "summary:" in response.lower()

agent = Agent(validator=contains_summary, validation_retries=2)

Format check

import re

def is_email_format(response: str) -> bool:
    pattern = r'^[\w\.-]+@[\w\.-]+\.\w+$'
    return bool(re.match(pattern, response.strip()))

agent = Agent(validator=is_email_format, validation_retries=2)

Combined with structured outputs

For strict type validation, use structured outputs instead:

from dataclasses import dataclass

@dataclass
class User:
    name: str
    age: int

# This guarantees the correct structure
result = await agent.run("Create a user", output=User)

Use validators for:

• Content validation (not just structure)
• Custom business rules
• Format requirements

Debug mode

See validation attempts with debug=True:

agent = Agent(
    validator=is_valid_json,
    validation_retries=3,
    debug=True,
)
# Prints: [Validation failed] Retrying...