smolagents

Integrate Agent Diff with HuggingFace smolagents

Overview

smolagents is HuggingFace’s lightweight agent framework. Agent Diff provides native tool support.

Installation

pip install agent-diff smolagents

Basic Integration

from agent_diff import AgentDiff, PythonExecutorProxy, create_smolagents_tool
from smolagents import CodeAgent, HfApiModel

# Initialize client and environment
client = AgentDiff()
env = client.init_env(
    templateService="slack",
    templateName="slack_default",
    impersonateUserId="U01AGENBOT9"
)

# Create executor and smolagents tool
python_executor = PythonExecutorProxy(env.environmentId, base_url=client.base_url)
python_tool = create_smolagents_tool(python_executor)

# Create agent
model = HfApiModel("Qwen/Qwen2.5-72B-Instruct")
agent = CodeAgent(
    tools=[python_tool],
    model=model,
    system_prompt="""You are a Slack assistant. Use the execute_python tool to 
    interact with Slack API at https://slack.com/api/*. 
    Authentication is handled automatically."""
)

# Start run and execute
run = client.start_run(envId=env.environmentId)
result = agent.run("Post 'Hello World!' to #general")

# Get diff
diff = client.diff_run(runId=run.runId)
print(diff.diff['inserts'])

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

With Different Models

smolagents supports multiple model backends:

from smolagents import HfApiModel, OpenAIServerModel

# HuggingFace Inference API
hf_model = HfApiModel("Qwen/Qwen2.5-72B-Instruct")

# OpenAI-compatible APIs
openai_model = OpenAIServerModel(
    model_id="gpt-4o",
    api_key="your-api-key"
)

# Use with agent
agent = CodeAgent(tools=[python_tool], model=openai_model)

Running Evaluations

from agent_diff import AgentDiff, PythonExecutorProxy, create_smolagents_tool
from smolagents import CodeAgent, HfApiModel

client = AgentDiff()
model = HfApiModel("Qwen/Qwen2.5-72B-Instruct")

# Get test suite
suites = client.list_test_suites(name="Slack Bench")
suite = client.get_test_suite(suites.testSuites[0].id, expand=True)

results = []

for test in suite.tests:
    env = client.init_env(testId=test.id)
    run = client.start_run(envId=env.environmentId, testId=test.id)
    
    # Create tool and agent
    executor = PythonExecutorProxy(env.environmentId, client.base_url)
    tool = create_smolagents_tool(executor)
    agent = CodeAgent(
        tools=[tool],
        model=model,
        system_prompt="Use execute_python to interact with Slack at https://slack.com/api/*"
    )
    
    # Run
    agent.run(test.prompt)
    
    # Evaluate
    result = client.evaluate_run(runId=run.runId)
    results.append({"test": test.name, "passed": result.passed})
    
    client.delete_env(envId=env.environmentId)

# Print results
passed = sum(1 for r in results if r['passed'])
print(f"\n{passed}/{len(results)} tests passed")

Next Steps

Code Executors

Learn how API interception works

Evaluations

Run benchmarks on your agents