Custom Evaluators

Build, register, and ship custom evaluators for Agent Control.

Creating Custom Evaluators

Partners and developers can create custom evaluators to extend Agent Control with their own detection capabilities. These evaluators can be published as wheels and installed in your Agent Control server. If you want to contribute an evaluator to the Agent Control repo, see Contributing Evaluator.

Evaluator Interface

Every evaluator implements the Evaluator base class:

from typing import Any

from agent_control_models import EvaluatorResult
from agent_control_evaluators import (
    Evaluator,
    EvaluatorConfig,
    EvaluatorMetadata,
    register_evaluator,
)


class MyEvaluatorConfig(EvaluatorConfig):
    """Configuration schema for your evaluator."""
    threshold: float = 0.5
    custom_option: str = "default"


@register_evaluator
class MyEvaluator(Evaluator[MyEvaluatorConfig]):
    """Your custom evaluator."""

    metadata = EvaluatorMetadata(
        name="my-evaluator",
        version="1.0.0",
        description="Detects custom patterns using proprietary logic",
        requires_api_key=True,  # Set to True if you need credentials
        timeout_ms=5000,
    )
    config_model = MyEvaluatorConfig

    def __init__(self, config: MyEvaluatorConfig) -> None:
        """Initialize with validated configuration."""
        super().__init__(config)
        # Set up any clients, load models, etc.

    async def evaluate(self, data: Any) -> EvaluatorResult:
        """
        Evaluate the input data.
        
        Args:
            data: The content to evaluate (string, dict, etc.)
            
        Returns:
            EvaluatorResult with:
              - matched: bool — Did this trigger the control?
              - confidence: float — How confident (0.0-1.0)?
              - message: str — Human-readable explanation
              - metadata: dict — Additional context for logging
        """
        # Your detection logic here
        score = await self._analyze(data)
        
        return EvaluatorResult(
            matched=score > self.config.threshold,
            confidence=score,
            message=f"Custom analysis score: {score:.2f}",
            metadata={
                "score": score,
                "threshold": self.config.threshold,
            }
        )
    
    async def _analyze(self, data: Any) -> float:
        """Your proprietary analysis logic."""
        # Call your API, run your model, etc.
        return 0.0

Evaluator Registration

Evaluators are discovered automatically via Python entry points. To make your evaluator available:

1. Create a Python package with your evaluator class decorated with @register_evaluator

2. Register as an entry point in your pyproject.toml:

   [project.entry-points."agent_control.evaluators"]
   my-evaluator = "my_package.evaluator:MyEvaluator"
   

3. Install it in the Agent Control environment

# Install your evaluator
pip install my-custom-evaluator

# It's now available

Optional Dependencies

If your evaluator has optional dependencies, override is_available():

try:
    import optional_dep
    AVAILABLE = True
except ImportError:
    AVAILABLE = False

@register_evaluator
class MyEvaluator(Evaluator[MyEvaluatorConfig]):
    @classmethod
    def is_available(cls) -> bool:
        return AVAILABLE

When is_available() returns False, the evaluator is silently skipped during registration.

Evaluator Best Practices

Practice	Why
Use Pydantic for config	Automatic validation and documentation
Implement timeouts	Prevent slow evaluators from blocking agents
Return confidence scores	Enable threshold-based filtering
Include metadata	Helps with debugging and observability
Handle errors gracefully	Respect the on_error configuration
Make API calls async	Don’t block the event loop

Example: Third-Party Integration

Here’s how a partner might integrate their content moderation API:

@register_evaluator
class ContentModerationEvaluator(Evaluator[ContentModerationEvaluatorConfig]):
    """Integration with Acme Content Moderation API."""
    
    metadata = EvaluatorMetadata(
        name="acme-content-mod",
        version="1.0.0",
        description="Acme Inc. content moderation",
        requires_api_key=True,
        timeout_ms=3000,
    )
    config_model = ContentModerationEvaluatorConfig

    def __init__(self, config: ContentModerationEvaluatorConfig) -> None:
        super().__init__(config)
        self.client = AcmeClient(api_key=os.getenv("ACME_API_KEY"))

    async def evaluate(self, data: Any) -> EvaluatorResult:
        result = await self.client.moderate(str(data))
        
        return EvaluatorResult(
            matched=result.flagged,
            confidence=result.confidence,
            message=result.reason,
            metadata={"categories": result.categories}
        )

Refer DeepEval Example for creating Custom Evaluators.