Foundry Code Interpreter Agent

Create an ad-hoc Foundry agent with Code Interpreter enabled using the new v2 Responses API.

The agent can write and execute Python code in a sandboxed environment to solve data analysis tasks, generate visualizations, and perform computations.

Prerequisites

The user must have completed the Foundry setup (the setup-foundry skill). They need:

•FOUNDRY_PROJECT_ENDPOINT -- the project endpoint URL
•FOUNDRY_MODEL_DEPLOYMENT_NAME -- the model deployment name (e.g. gpt-4.1-mini)

If these are not set, ask the user for the values.

Steps

1. Determine the Task

Ask the user what they want the code interpreter agent to do. Examples:

•"Analyze this CSV file and create a bar chart"
•"Solve this equation: sin(x) + x^2 = 42"
•"Generate a histogram of this data"
•"Calculate statistics for this dataset"

If the user provides a file, note the file path for upload.

2. Run the Code Interpreter Agent

Use the following Python script. Replace placeholders with actual values.

Without file upload (computation / code generation):

bash

python3 << 'PYEOF'
import os
from azure.identity import DefaultAzureCredential
from azure.ai.projects import AIProjectClient
from azure.ai.projects.models import PromptAgentDefinition, CodeInterpreterTool, CodeInterpreterToolAuto

endpoint = os.environ.get("FOUNDRY_PROJECT_ENDPOINT", "<ENDPOINT>")
model = os.environ.get("FOUNDRY_MODEL_DEPLOYMENT_NAME", "<MODEL>")

project_client = AIProjectClient(
    endpoint=endpoint,
    credential=DefaultAzureCredential(),
)

with project_client:
    openai_client = project_client.get_openai_client()

    # Create agent with code interpreter
    agent = project_client.agents.create_version(
        agent_name="CodeInterpreterAgent",
        definition=PromptAgentDefinition(
            model=model,
            instructions="You are a data analysis and computation assistant. Write and execute Python code to solve the user's problem. Show results clearly.",
            tools=[CodeInterpreterTool(container=CodeInterpreterToolAuto())],
        ),
        description="Ad-hoc code interpreter agent.",
    )
    print(f"Agent created: {agent.name} v{agent.version}")

    # Create a conversation
    conversation = openai_client.conversations.create()
    print(f"Conversation: {conversation.id}")

    # Send the user's request
    response = openai_client.responses.create(
        conversation=conversation.id,
        input="USER_PROMPT_HERE",
        extra_body={"agent": {"name": agent.name, "type": "agent_reference"}},
    )

    # Print the response
    for item in response.output:
        if item.type == "message":
            for content in item.content:
                if content.type == "output_text":
                    print(content.text)

    # Clean up
    project_client.agents.delete_version(agent_name=agent.name, agent_version=agent.version)
    print("Agent cleaned up.")
PYEOF

With file upload (data analysis):

bash

python3 << 'PYEOF'
import os
from azure.identity import DefaultAzureCredential
from azure.ai.projects import AIProjectClient
from azure.ai.projects.models import PromptAgentDefinition, CodeInterpreterTool, CodeInterpreterToolAuto

endpoint = os.environ.get("FOUNDRY_PROJECT_ENDPOINT", "<ENDPOINT>")
model = os.environ.get("FOUNDRY_MODEL_DEPLOYMENT_NAME", "<MODEL>")
file_path = "FILE_PATH_HERE"

project_client = AIProjectClient(
    endpoint=endpoint,
    credential=DefaultAzureCredential(),
)

with project_client:
    openai_client = project_client.get_openai_client()

    # Upload the file
    with open(file_path, "rb") as f:
        uploaded_file = openai_client.files.create(purpose="assistants", file=f)
    print(f"File uploaded: {uploaded_file.id}")

    # Create agent with code interpreter and file access
    agent = project_client.agents.create_version(
        agent_name="DataAnalysisAgent",
        definition=PromptAgentDefinition(
            model=model,
            instructions="You are a data analysis assistant. Analyze the uploaded file and fulfil the user's request. Generate charts or output files when appropriate.",
            tools=[CodeInterpreterTool(container=CodeInterpreterToolAuto(file_ids=[uploaded_file.id]))],
        ),
        description="Ad-hoc data analysis agent with file.",
    )
    print(f"Agent created: {agent.name} v{agent.version}")

    # Create a conversation
    conversation = openai_client.conversations.create()

    # Send the user's request
    response = openai_client.responses.create(
        conversation=conversation.id,
        input="USER_PROMPT_HERE",
        extra_body={"agent": {"name": agent.name, "type": "agent_reference"}},
    )

    # Print response and download generated files
    for item in response.output:
        if item.type == "message":
            for content in item.content:
                if content.type == "output_text":
                    print(content.text)
                    # Check for file citations
                    if hasattr(content, "annotations") and content.annotations:
                        for ann in content.annotations:
                            if ann.type == "container_file_citation":
                                file_content = openai_client.containers.files.content.retrieve(
                                    file_id=ann.file_id, container_id=ann.container_id
                                )
                                safe_name = os.path.basename(ann.filename)
                                with open(safe_name, "wb") as out:
                                    out.write(file_content.read())
                                print(f"Downloaded: {safe_name}")

    # Clean up
    project_client.agents.delete_version(agent_name=agent.name, agent_version=agent.version)
    print("Agent cleaned up.")
PYEOF

3. Present Results

After running the script:

•Show the text output from the agent
•If files were generated (charts, CSVs), tell the user where they were saved
•If there were errors, explain what went wrong and suggest fixes

4. Follow-up

The conversation is stateful. If the user wants to continue the analysis, reuse the same conversation ID and agent. Only clean up when the user is done.

Supported File Types

The code interpreter supports: .csv, .json, .xlsx, .txt, .pdf, .py, .md, .html, .png, .jpg, .gif, and more.

Notes

•Code Interpreter runs Python in a sandboxed container with common data science packages (pandas, matplotlib, numpy, etc.)
•Each session is active for 1 hour with a 30-minute idle timeout
•Additional charges apply beyond standard token-based fees
•For custom packages, a custom code interpreter container can be configured