Gestion des demandes d’outils d’utilisation d’ordinateur émanant des agents pendant les conversations

Lorsque votre agent demande un outil, la réponse à votre opération d’API InvokeAgent inclut des données utiles returnControl qui comprennent l’outil à utiliser et l’action de l’outil dans les invocationInputs. Pour plus d’informations sur le contrôle des retours au développeur de l’agent, consultez Redonner le contrôle au développeur de l’agent en envoyant les informations obtenues dans une réponse InvokeAgent.

Renvoyer un exemple de contrôle

Voici un exemple de données utiles returnControl avec une demande d’utilisation de l’outil ANTHROPIC.Computer avec l’action screenshot.

{
    "returnControl": {
        "invocationId": "invocationIdExample",
        "invocationInputs": [{
            "functionInvocationInput": {
                "actionGroup": "my_computer",
                "actionInvocationType": "RESULT",
                "agentId": "agentIdExample",
                "function": "computer",
                "parameters": [{
                    "name": "action",
                    "type": "string",
                    "value": "screenshot"
                }]
            }
        }]
    }
}

Exemple de code pour analyser la demande d’outil

Le code suivant montre comment extraire le choix de l’outil informatique dans une réponse InvokeAgent, le mapper à des implémentations d’outils fictifs pour différents outils, puis envoyer le résultat de l’utilisation de l’outil dans une demande InvokeAgent suivante.

Pour plus d’informations sur les outils d’utilisation d’ordinateur, notamment des exemples de mise en œuvre et des descriptions d’outils, consultez Utilisation d’ordinateur (version bêta) dans la documentation Anthropic.

import boto3
from botocore.exceptions import ClientError
import json


def handle_computer_action(action_params):
    """
    Maps computer actions, like taking screenshots and moving the mouse to mock implementations and returns
    the result.

    Args:
        action_params (dict): Dictionary containing the action parameters
            Keys:
                - action (str, required): The type of action to perform (for example 'screenshot' or 'mouse_move')
                - coordinate (str, optional): JSON string containing [x,y] coordinates for mouse_move

    Returns:
        dict: Response containing the action result.
    """

    action = action_params.get('action')
    if action == 'screenshot':
        # Mock screenshot response
        with open("mock_screenshot.png", 'rb') as image_file:
            image_bytes = image_file.read()
        return {
            "IMAGES": {
                "images": [
                    {
                        "format": "png",
                        "source": {
                            "bytes": image_bytes
                        },
                    }
                ]
            }
        }
    elif action == 'mouse_move':
        # Mock mouse movement
        coordinate = json.loads(action_params.get('coordinate', '[0, 0]'))
        return {
            "TEXT": {
                "body": f"Mouse moved to coordinates {coordinate}"
            }
        }
    elif action == 'left_click':
        # Mock mouse left click
        return {
            "TEXT": {
                "body": f"Mouse left clicked"
            }
        }
    elif action == 'right_click':
        # Mock mouse right click
        return {
            "TEXT": {
                "body": f"Mouse right clicked"
            }
        }

    ### handle additional actions here


def manage_computer_interaction(bedrock_agent_runtime_client, agent_id, alias_id):
    """
    Manages interaction between an Amazon Bedrock agent and computer use functions.

    Args:
        bedrock_agent_runtime_client: Boto3 client for Bedrock agent runtime
        agent_id (str): The ID of the agent
        alias_id (str): The Alias ID of the agent

    The function:
    - Initiates a session with initial prompt
    - Makes agent requests with appropriate parameters
    - Processes response chunks and return control events
    - Handles computer actions via handle_computer_action()
    - Continues interaction until task completion
    """
    session_id = "session123"
    initial_prompt = "Open a browser and go to a website"
    computer_use_results = None
    current_prompt = initial_prompt

    while True:
        # Make agent request with appropriate parameters
        invoke_params = {
            "agentId": agent_id,
            "sessionId": session_id,
            "inputText": current_prompt,
            "agentAliasId": alias_id,
        }

        # Include session state if we have results from previous iteration
        if computer_use_results:
            invoke_params["sessionState"] = computer_use_results["sessionState"]

        try:
            response = bedrock_agent_runtime_client.invoke_agent(**invoke_params)
        except ClientError as e:
            print(f"Error: {e}")

        has_return_control = False

        # Process the response
        for event in response.get('completion'):
            if 'chunk' in event:
                chunk_content = event['chunk'].get('bytes', b'').decode('utf-8')
                if chunk_content:
                    print("\nAgent:", chunk_content)

            if 'returnControl' in event:
                has_return_control = True
                invocationId = event["returnControl"]["invocationId"]
                if "invocationInputs" in event["returnControl"]:
                    for invocationInput in event["returnControl"]["invocationInputs"]:
                        func_input = invocationInput["functionInvocationInput"]

                        # Extract action parameters
                        params = {p['name']: p['value'] for p in func_input['parameters']}

                        # Handle computer action and get result
                        action_result = handle_computer_action(params)

                        # Print action result for testing
                        print("\nExecuting function:", func_input['function'])
                        print("Parameters:", params)

                        # Prepare the session state for the next request
                        computer_use_results = {
                            "sessionState": {
                                "invocationId": invocationId,
                                "returnControlInvocationResults": [{
                                    "functionResult": {
                                        "actionGroup": func_input['actionGroup'],
                                        "responseState": "REPROMPT",
                                        "agentId": func_input['agentId'],
                                        "function": func_input['function'],
                                        "responseBody": action_result
                                    }
                                }]
                            }
                        }

        # If there's no return control event, the task is complete
        if not has_return_control:
            print("\nTask completed!")
            break

        # Use empty string as prompt for subsequent iterations
        current_prompt = ""
def main():
    bedrock_agent_runtime_client = boto3.client(service_name="bedrock-agent-runtime",
                                         region_name="REGION"
                                         )

    agent_id = "AGENT_ID"
    alias_id = "ALIAS_ID"

    manage_computer_interaction(bedrock_agent_runtime_client, agent_id, alias_id)


if __name__ == "__main__":
    main()

La sortie doit ressembler à ce qui suit :

Executing function: computer
Parameters: {'action': 'screenshot'}

Executing function: computer
Parameters: {'coordinate': '[467, 842]', 'action': 'mouse_move'}

Executing function: computer
Parameters: {'action': 'left_click'}

Agent: I've opened Firefox browser. Which website would you like to visit?

Task completed!