Edit

Share via

Run code by using code interpreter in Azure SRE Agent

The SRE Agent code interpreter enables you to execute Python code and shell commands in a secure, isolated sandbox environment. Use Code Interpreter to analyze data, create visualizations, generate PDF reports, and automate file operations without leaving your SRE Agent conversation.

In this article, you learn how to:

  • Execute Python code to analyze data and create visualizations
  • Run shell commands for file operations
  • Generate and download PDF reports
  • Work with files in the sandbox environment

Prerequisites

To create an agent, you need to grant your agent the correct permissions, configure the correct settings, and grant access to the right resources:

  • Azure account: You need an Azure account with an active subscription. If you don't already have one, you can create an account for free.

  • Security context: Make sure that your user account has the Microsoft.Authorization/roleAssignments/write permissions as either Role Based Access Control Administrator or User Access Administrator.

  • Firewall settings: Add *.azuresre.ai to the allowlist in your firewall settings. Some networking profiles might block access to *.azuresre.ai domain by default.

How code interpreter works

The SRE Agent code interpreter runs in an isolated Azure Container Apps session with the following characteristics:

  • No network access: The sandbox can't make outbound HTTP or HTTPS requests.
  • No process spawning: Commands like subprocess and os.system are blocked.
  • No package installation: pip install and conda install aren't available.
  • Isolated file system: All files must be saved to /mnt/data/.

These restrictions ensure that code execution is secure and predictable. Common data science libraries like pandas, matplotlib, and seaborn are preinstalled.

Get started with code interpreter

Code interpreter tools are automatically available in your SRE Agent conversations. Ask the agent to perform tasks by using natural language.

Example prompts

Try these prompts to get started:

  • "Analyze this CSV data and create a chart showing incidents by category."
  • "Generate a PDF summary of today's incidents."
  • "Parse this JSON file and extract the error patterns."
  • "List all files in the session and show their sizes."

Your first Python script

The following example demonstrates how to create a visualization:

import matplotlib.pyplot as plt
import numpy as np

# Generate sample data
categories = ['Critical', 'High', 'Medium', 'Low']
counts = [5, 12, 28, 45]

# Create bar chart
plt.figure(figsize=(10, 6))
plt.bar(categories, counts, color=['red', 'orange', 'yellow', 'green'])
plt.title('Incidents by Severity')
plt.xlabel('Severity Level')
plt.ylabel('Count')
plt.savefig('/mnt/data/incidents_by_severity.png', dpi=150)

When the code runs successfully, the agent returns the image inline in the conversation.

Scenarios

Analyze incident data

Use code interpreter to transform raw incident data into actionable insights and visualizations.

  1. Ask SRE Agent to create a CSV file of all the incidents that have occurred within the last month using the following prompt:

    Create a CSV file of all the incidents that 
occurred in the last month.

Name the file `incidents.csv` and save it
to `/mnt/data`.

  2. Use Python to analyze incident patterns and identify trends:

    import pandas as pd

# Load incident data
df = pd.read_csv('/mnt/data/incidents.csv')

# Calculate summary statistics
summary = df.groupby('category').agg({
    'id': 'count',
    'resolution_time': 'mean'
}).rename(columns={'id': 'count', 'resolution_time': 'avg_resolution_hours'})

# Export results
summary.to_csv('/mnt/data/incident_summary.csv')
print(summary)

Create visualizations

Generate charts to communicate insights.

import matplotlib.pyplot as plt
import seaborn as sns

# Load data
df = pd.read_csv('/mnt/data/incidents.csv')

# Create visualization
plt.figure(figsize=(12, 6))
sns.barplot(data=df, x='category', y='count', palette='viridis')
plt.title('Incidents by Category')
plt.xticks(rotation=45)
plt.tight_layout()
plt.savefig('/mnt/data/category_chart.png', dpi=150)

Generate PDF reports

Create formatted PDF documents for stakeholders.

from reportlab.pdfgen import canvas
from reportlab.lib.pagesizes import letter

# Create PDF
c = canvas.Canvas('/mnt/data/weekly_report.pdf', pagesize=letter)

# Add content
c.setFont('Helvetica-Bold', 18)
c.drawString(72, 750, 'Weekly Incident Report')

c.setFont('Helvetica', 12)
c.drawString(72, 720, 'Period: January 10-17, 2026')
c.drawString(72, 700, 'Total incidents: 90')
c.drawString(72, 680, 'Critical incidents: 5')
c.drawString(72, 660, 'Average resolution time: 4.2 hours')

c.save()

The agent returns a download link for the generated PDF.

Process log files

Search and analyze log files by using shell commands.

# Count error occurrences by type
grep -E "ERROR|WARN|FATAL" application.log | sort | uniq -c | sort -rn

# Find the most recent log entries
tail -100 application.log

Tool reference

Code Interpreter provides the following tools:

ExecutePythonCode

Runs Python code in an isolated sandbox.

Parameter Type Required Description
pythonCode string Yes Python code to run (up to 20,000 characters)
timeoutSeconds integer No Execution timeout in seconds (default: 120, maximum: 900)

Output: Images show inline as markdown. Data files return as download links.

GeneratePdfReport

Creates PDF documents from Python code.

Parameter Type Required Description
pythonCode string Yes Python code that creates a PDF file
expectedOutputFilename string Yes Path where the PDF is saved (for example, report.pdf)
saveAsFilename string Yes Download filename (for example, daily_summary.pdf)
timeoutSeconds integer No Execution timeout in seconds (default: 180, maximum: 900)

RunShellCommand

Runs shell commands in the sandbox.

Parameter Type Required Description
command string Yes Shell command to run
explanation string No Description to log with the command
isBackground boolean No Must be false (background jobs aren't supported)
timeoutSeconds integer No Execution timeout (default: 120, max: 240)

Note

The commands run from /mnt/data/. Use relative paths and chain commands by using ; instead of &&.

File operations

Tool Description
ReadSessionFile Reads the contents of a file from /mnt/data/.
SearchSessionFiles Searches for text within files using grep-style patterns.
ListSessionFiles Lists all files in the session with metadata.
GetSessionFile Downloads a file from the session.
UploadFileToSession Uploads a file for further processing.

Supported file types

  • Images: PNG, JPG, GIF, SVG, WebP, BMP, TIFF, EPS
  • Data: CSV, TSV, Excel, JSON, XML, YAML, HDF5, NetCDF, pickle
  • Documents: PDF, HTML, Markdown, Office formats
  • Code: Python, Jupyter notebooks, R, SQL
  • Archives: ZIP, TAR, GZ

Limitations

For security reasons, the system blocks the following operations:

Category Blocked operations
Process spawning subprocess, os.system, os.popen, os.spawn*
Network access Outbound HTTP/HTTPS requests
Package installation pip install, conda install
File system Access outside /mnt/data/

Troubleshoot code interpreter

Code execution fails

  • Verify your code doesn't use blocked operations like subprocess or network calls.
  • Check that all file paths point to /mnt/data/.
  • Ensure your code is under 20,000 characters.

Files don't appear in output

  • Confirm you save files to /mnt/data/.
  • Use ListSessionFiles to verify file creation.
  • Check that the file extension is supported.

Timeout errors

  • Increase the timeoutSeconds parameter (maximum 900 seconds).
  • Optimize your code to reduce execution time.
  • Split large operations into smaller chunks.

Related content

  • Last updated on