agent-diff 0.1.2

Python SDK for Agent Diff - test AI agents and train models against replicas of services

Agent Diff Python SDK

Python SDK for testing AI agents against isolated replicas of production services.

Installation

uv add agent-diff

Quick Start

from agent_diff import AgentDiff
from agent_diff import PythonExecutorProxy, create_openai_tool

# Self-hosted (defaults to http://localhost:8000)
client = AgentDiff()

# With authentication 
client = AgentDiff(
    api_key="your-api-key",
    base_url="https://your-instance.com"
)

# 1. Create an isolated environment
env = client.init_env(
    templateService="slack",
    templateName="slack_default",
    impersonateUserId="U123456",
    ttlSeconds=1800
)

# 2. Create executor with automatic API interception

python_executor = PythonExecutorProxy(env.environmentId, base_url=client.base_url)
python_tool = create_openai_tool(python_executor)

# 3. Take before snapshot
run = client.start_run(envId=env.environmentId)

# 4. Run your agent (API calls are automatically intercepted)
from agents import Agent

agent = Agent(model="gpt-4o", tools=[python_tool])
response = agent.run("Send a message to #general saying 'Hello!'")

# 5. Compute the diff
diff = client.diff_run(runId=run.runId)

# Inspect changes
diff.diff['inserts']   # New records
diff.diff['updates']   # Modified records
diff.diff['deletes']   # Deleted records

# 6. Cleanup
client.delete_env(envId=env.environmentId)

Code Execution Proxies

Agent Diff provides code execution proxies that automatically intercept API calls and route them to isolated test environments. This enables agents with code execution capabilities to interact with service replicas without any code changes.

How It Works

When your agent executes Python or Bash code:

1. The executor wraps your code with interception logic
2. API calls to https://api.slack.com → http://localhost:8000/api/env/{env_id}/services/slack
3. API calls to https://api.linear.app → http://localhost:8000/api/env/{env_id}/services/linear
4. Your agent sees real API responses from the isolated environment

Available Executors

PythonExecutorProxy

Intercepts Python requests and urllib libraries:

from agent_diff import PythonExecutorProxy, create_openai_tool

python_executor = PythonExecutorProxy(env.environmentId, base_url=client.base_url)
python_tool = create_openai_tool(python_executor)

# Works with OpenAI Agents SDK, LangChain, smolagents
agent = Agent(model="gpt-5", tools=[python_tool])
agent.run("Send a Slack message to #general")

BashExecutorProxy

Intercepts curl commands:

from agent_diff import BashExecutorProxy, create_openai_tool

bash_executor = BashExecutorProxy(env.environmentId, base_url=client.base_url)
bash_tool = create_openai_tool(bash_executor)

agent = Agent(model="gpt-5", tools=[bash_tool])
agent.run("Use curl to post a message to Slack")

Framework Support

Create tools for popular agent frameworks:

from agent_diff import create_openai_tool, create_langchain_tool, create_smolagents_tool

# OpenAI Agents SDK
openai_tool = create_openai_tool(python_executor)

# LangChain
langchain_tool = create_langchain_tool(python_executor)

# HuggingFace smolagents
smolagents_tool = create_smolagents_tool(python_executor)

Direct Execution

For custom frameworks or direct usage:

python_executor = PythonExecutorProxy(env.environmentId, base_url=client.base_url)

result = python_executor.execute("""
import requests
response = requests.post('https://api.slack.com/api/chat.postMessage', json={
    'channel': '#general',
    'text': 'Hello from Agent Diff!'
})
print(response.json())
""")

if result["status"] == "success":
    print(result["stdout"])
else:
    print(result["stderr"])

Environments

Create isolated, ephemeral replicas of services:

env = client.init_env(
    templateService="slack",
    templateName="slack_default",
    impersonateUserId="U123",
    ttlSeconds=3600
)

# Access environment details
env.environmentId
env.environmentUrl
env.expiresAt

# Delete when done
client.delete_env(env.environmentId)

Test Suites

To run evaluations:

suite = client.get_test_suite("slack-bench")
# Returns: {"tests": [{"id": "...", "prompt": "Send hello to #general"}, ...]}


evaluation_results = []

for test in suite['tests']:
    prompt = test['prompt']
    test_id = test['id']

    env = client.init_env(testId=test_id)
    run = client.start_run(envId=env.environmentId, testId=test_id)

    # Create executor with automatic API interception
    python_executor = PythonExecutorProxy(env.environmentId, base_url=client.base_url)
    python_tool = create_openai_tool(python_executor)

    # Run your agent with the tool
    agent = Agent(model="gpt-5", tools=[python_tool])
    response = agent.run(prompt)

    evaluation_result = client.evaluate_run(run.runId)  # Returns score, runId, status and Score (0/1)

    evaluation_results.append(evaluation_result)

    client.delete_env(envId=env.environmentId)

Templates

List and create environment templates:

# List available templates
templates = client.list_templates()

# Create custom template - you can populate the replica and turn it into a template with custom data
custom = client.create_template_from_environment(
    environmentId=env.environmentId,
    service="slack",
    name="my_template",
    description="Custom template",
    visibility="private"  # "private" means only you can view the template
)

License

MIT License - see LICENSE file for details.