Criar avaliador

agentcore add evaluator \
  --name "your_custom_evaluator_name" \
  --config custom_evaluator_config.json \
  --level "TRACE"

Esse comando adiciona o avaliador à sua agentcore.json configuração local. Corra agentcore deploy para criá-lo em sua AWS conta.

Insira um nome para seu avaliador personalizado.

Selecione o nível de avaliação: Session, Trace ou Tool Call.

Escolha o modelo de juiz LLM para avaliação.

Insira suas instruções de avaliação. O prompt deve incluir pelo menos um espaço reservado: {context} para o histórico de conversas ou {available_tools} para a lista de ferramentas.

Selecione uma predefinição de escala de avaliação ou defina uma escala personalizada.

Revise a configuração do avaliador e pressione Enter para confirmar.

import json
from bedrock_agentcore_starter_toolkit import Evaluation

eval_client = Evaluation()

# Load the configuration JSON file
with open('custom_evaluator_config.json') as f:
    evaluator_config = json.load(f)

# Create the custom evaluator
custom_evaluator = eval_client.create_evaluator(
    name="your_custom_evaluator_name",
    level="TRACE",
    description="Response quality evaluator",
    config=evaluator_config
)

import boto3
import json

client = boto3.client('bedrock-agentcore-control')

# Load the configuration JSON file
with open('custom_evaluator_config.json') as f:
    evaluator_config = json.load(f)

# Create the custom evaluator
response = client.create_evaluator(
    evaluatorName="your_custom_evaluator_name",
    level="TRACE",
    evaluatorConfig=evaluator_config
)

aws bedrock-agentcore-control create-evaluator \
    --evaluator-name 'your_custom_evaluator_name' \
    --level TRACE \
    --evaluator-config file://custom_evaluator_config.json

Esse avaliador usa um LLM para comparar as trajetórias esperadas e reais da ferramenta, permitindo um julgamento diferenciado — por exemplo, tolerando pequenos desvios, como chamadas extras de ferramentas auxiliares. Ele usa os expected_tool_trajectory actual_tool_trajectory marcadores de posição e.

Salve o seguinte comotrajectory_compliance_config.json:

{
  "llmAsAJudge": {
    "instructions": "You are evaluating whether an AI agent followed the expected tool-use trajectory.\n\nExpected trajectory (ordered list of tool names):\n{expected_tool_trajectory}\n\nActual trajectory (ordered list of tool names the agent used):\n{actual_tool_trajectory}\n\nFull session context:\n{context}\n\nAvailable tools:\n{available_tools}\n\nCompare the expected and actual trajectories. Consider whether the agent called the right tools in the right order. Minor deviations (e.g., an extra logging tool call) are acceptable if the core trajectory is preserved.",
    "ratingScale": {
      "numerical": [
        { "label": "No Match",      "value": 0.0, "definition": "The actual trajectory has no meaningful overlap with the expected trajectory" },
        { "label": "Partial Match", "value": 0.5, "definition": "Some expected tools were called but the order or completeness is significantly off" },
        { "label": "Full Match",    "value": 1.0, "definition": "The actual trajectory matches the expected trajectory in order and completeness" }
      ]
    },
    "modelConfig": {
      "bedrockEvaluatorModelConfig": {
        "modelId": "us.anthropic.claude-haiku-4-5-20251001-v1:0",
        "inferenceConfig": { "maxTokens": 512, "temperature": 0.0 }
      }
    }
  }
}

Crie o avaliador:

aws bedrock-agentcore-control create-evaluator \
  --evaluator-name 'TrajectoryCompliance' \
  --level SESSION \
  --description 'Evaluates whether the agent followed the expected tool trajectory.' \
  --evaluator-config file://trajectory_compliance_config.json

Esse avaliador verifica se o comportamento do agente satisfaz um conjunto de afirmações, retornando um veredicto categórico. PASS/FAIL/INCONCLUSIVE Ele usa o assertions espaço reservado junto com context e. available_tools

Salve o seguinte comoassertion_checker_config.json:

{
  "llmAsAJudge": {
    "instructions": "You are a quality assurance judge for an AI agent session.\n\nSession context (full conversation history):\n{context}\n\nAvailable tools:\n{available_tools}\n\nAssertions to verify:\n{assertions}\n\nFor each assertion, determine if the session satisfies it. The overall verdict should be PASS only if ALL assertions are satisfied. If any assertion fails, the verdict is FAIL. If the session data is insufficient to determine, verdict is INCONCLUSIVE.",
    "ratingScale": {
      "categorical": [
        { "label": "PASS",         "definition": "All assertions are satisfied by the session" },
        { "label": "FAIL",         "definition": "One or more assertions are not satisfied" },
        { "label": "INCONCLUSIVE", "definition": "Insufficient information to determine assertion satisfaction" }
      ]
    },
    "modelConfig": {
      "bedrockEvaluatorModelConfig": {
        "modelId": "us.anthropic.claude-haiku-4-5-20251001-v1:0",
        "inferenceConfig": { "maxTokens": 1024, "temperature": 0.0 }
      }
    }
  }
}

Crie o avaliador:

aws bedrock-agentcore-control create-evaluator \
  --evaluator-name 'AssertionChecker' \
  --level SESSION \
  --description 'Checks whether the agent session satisfies a set of assertions.' \
  --evaluator-config file://assertion_checker_config.json

Esse avaliador compara a resposta real do agente com uma resposta esperada, pontuando a similaridade semântica. Ele usa o expected_response espaço reservado para receber a verdade fundamental no momento da avaliação.

Salve o seguinte comoresponse_similarity_config.json:

{
  "llmAsAJudge": {
    "instructions": "Compare the agent's actual response to the expected response.\n\nConversation context:\n{context}\n\nAgent's actual response:\n{assistant_turn}\n\nExpected response:\n{expected_response}\n\nEvaluate semantic similarity. The agent does not need to match word-for-word, but the meaning, key facts, and intent should align. Penalize missing critical information or contradictions.",
    "ratingScale": {
      "numerical": [
        { "label": "No Match",        "value": 0.0,  "definition": "The response contradicts or is completely unrelated to the expected response" },
        { "label": "Low Similarity",   "value": 0.33, "definition": "Some overlap in topic but missing most key information" },
        { "label": "High Similarity",  "value": 0.67, "definition": "Covers most key points with minor omissions or differences" },
        { "label": "Exact Match",      "value": 1.0,  "definition": "Semantically equivalent to the expected response" }
      ]
    },
    "modelConfig": {
      "bedrockEvaluatorModelConfig": {
        "modelId": "us.anthropic.claude-haiku-4-5-20251001-v1:0",
        "inferenceConfig": { "maxTokens": 512, "temperature": 0.0 }
      }
    }
  }
}

Crie o avaliador:

aws bedrock-agentcore-control create-evaluator \
  --evaluator-name 'ResponseSimilarity' \
  --level TRACE \
  --description 'Evaluates how closely the agent response matches the expected response.' \
  --evaluator-config file://response_similarity_config.json