Custom Evaluator Agent-as-Judge Evaluation

Demonstrates using a custom evaluator agent for judging.

With a custom evaluator, place the criterion and rubric in that agent’s instructions. The outer criteria and additional_guidelines are not automatically added to its prompt. The framework supplies the scoring output schema and applies the threshold to the returned score.

agent_as_judge_custom_evaluator.py

"""
Custom Evaluator Agent-as-Judge Evaluation
==========================================

Demonstrates using a custom evaluator agent for judging.
"""

from agno.agent import Agent
from agno.eval.agent_as_judge import AgentAsJudgeEval
from agno.models.openai import OpenAIChat

# ---------------------------------------------------------------------------
# Create Agent
# ---------------------------------------------------------------------------
agent = Agent(
    model=OpenAIChat(id="gpt-5.6-luna"),
    instructions="Explain technical concepts simply.",
)

# ---------------------------------------------------------------------------
# Create Evaluator Agent
# ---------------------------------------------------------------------------
custom_evaluator = Agent(
    model=OpenAIChat(id="gpt-5.6-luna"),
    description="Strict technical evaluator",
    instructions="You are a strict evaluator. Only give high scores to exceptionally clear and accurate explanations.",
)

# ---------------------------------------------------------------------------
# Create Evaluation
# ---------------------------------------------------------------------------
evaluation = AgentAsJudgeEval(
    name="Technical Accuracy",
    criteria="Explanation must be technically accurate and comprehensive",
    scoring_strategy="numeric",
    threshold=8,
    evaluator_agent=custom_evaluator,
)

# ---------------------------------------------------------------------------
# Run Evaluation
# ---------------------------------------------------------------------------
if __name__ == "__main__":
    response = agent.run("What is machine learning?")
    result = evaluation.run(
        input="What is machine learning?",
        output=str(response.content),
        print_results=True,
    )
    print(f"Score: {result.results[0].score}/10")
    print(f"Passed: {result.results[0].passed}")

Run the Example

Set up your virtual environment

uv venv --python 3.12
source .venv/bin/activate

Install dependencies

uv pip install -U agno openai

Export your OpenAI API key

export OPENAI_API_KEY="your_openai_api_key_here"

Give the custom judge its criterion

Immediately after creating custom_evaluator and before constructing evaluation, set:

custom judge instructions

custom_evaluator.instructions = [
    "You are a strict technical evaluator.",
    "Explanation must be technically accurate and comprehensive",
    "Only give high scores to exceptionally clear and accurate explanations.",
    "Return an integer score from 1 to 10 and explain the score: "
    "1 is incorrect or irrelevant, 5 is partly correct with material gaps, "
    "8 is accurate and comprehensive, and 10 is exceptionally clear and complete.",
]

Run the example

Save the code above as agent_as_judge_custom_evaluator.py, then run:

python agent_as_judge_custom_evaluator.py

Full source: cookbook/09_evals/agent_as_judge/agent_as_judge_custom_evaluator.py