[Feature Request] Add integration to smolagents #398
Description
Proposal
Hello OpenPipe Team! Thanks for creating such a cool project for agent RL.
I recently saw the LangGraph integration announcement and tested it, and it works great :)
I'd like to propose adding Smolagents integration as well.
I believe Smolagents is fit for ART for several reasons:
- Simplicity: Smolagents follows a philosophy of simplicity, making it accessible while maintaining capabilities
- Community: Like LangGraph, it provides diverse community examples and recipes. Particularly, recipes like open_deep_research are ideal use cases where ART's RL techniques can provide significant value
- Integration-friendly: Smolagents' class architecture (ReACT, Memory, Message, Logging, etc.) is well-structured and easily accessible, making it straightforward to integrate with ART
Example
from art.smolagents import init_chat_model
from smolagents import CodeAgent, WebSearchTool
import art
import weave
@weave.op
async def rollout(model: art.Model, scenario: ResearchScenario) -> art.Trajectory:
chat_model = init_chat_model(model.name, temperature=1.0)
# Define available tools
tools = [WebSearchTool()]
# Create the Smolagents ReAct agent
agent = CodeAgent(tools=tools, model=chat_model, stream_outputs=True)
# Execute workflow (need to verify async support in smolagents)
return await agent.run(scenario.messages)
from art.smolagents import wrap_rollout
...
for batch in training_iterator:
groups = []
for scenario in batch.items:
groups.append(
art.TrajectoryGroup([
wrap_rollout(rollout)(model, scenario)
for _ in range(training_config["rollouts_per_group"])
])
)
...Roadmap
- Add
art.smolagentsintegration module with wrapper functions - Add documentation and examples.
I'd be happy to work on implementing this feature if the team is interested.
Would love to hear your thoughts and any ideas! Thanks 😊