What do you like best about Hugging Face smolagents?
Three things stood out:
1. CodeAgent over JSON tool-calling
Most agent frameworks force the LLM to output structured JSON dictionaries. smolagents lets the LLM write actual Python code instead. This sounds scary but it's the whole point — it uses 30% fewer LLM steps and the agent can do things like loop over search results, transform data, and call tools in a single step instead of going back-and-forth 5 times.
In Demo 2, the agent couldn't just call a tool — it noticed an error, imported HfApi directly, called it with different arguments, compared results across two categories, and synthesized a final answer. All in one step.
2. The @tool decorator
@tool
def get_most_popular_model(task: str) -> str:
"""Returns the most downloaded model for a given task on HF Hub."""
model = next(iter(list_models(filter=task, sort="downloads", direction=-1)))
return model.id
That's it. One function + docstring = working tool. No YAML schemas, no register_tool() calls, no boilerplate. The docstring becomes the LLM's instruction manual automatically.
3. Model-agnostic flexibility
We swapped from HF Inference API to an Unsloth GGUF endpoint mid-POC with a one-line model change. No framework lock-in. InferenceClientModel, OpenAIModel, LiteLLMModel, TransformersModel — pick whichever fits your stack. Most agent frameworks are married to one or two providers. Review collected by and hosted on G2.com.
What do you dislike about Hugging Face smolagents?
Four main pain points from the POC:
1. Zero visibility into cost/token burn
We had no idea how many tokens we were consuming until a run finished (or timed out). In Demo 2, the agent burned 32,000+ tokens across 8 steps just to answer a simple comparison question. There's no running counter, no budget cap, no warning when you're about to blow through credits. For production use, this is a dealbreaker — you need to know your burn rate.
2. Built-in tools can silently break
Demo 2 used @tool with list_models(filter=task, sort="downloads", direction=-1). That tool failed at runtime because huggingface_hub removed the `direction` parameter. The library's own default tools are not version-pinned or tested against the latest API. The agent recovered autonomously (which was impressive), but the bug shouldn't have been there in the first place.
3. Multi-agent has no safety nets
Demo 3 (manager + web search agent) ran for 15+ minutes without producing output. No timeout, no max-iterations per sub-agent, no way to cancel cleanly. In a production pipeline, a runaway multi-agent loop could burn hundreds of thousands of tokens before you notice. You have to build all the guardrails yourself.
4. Error messages are raw tracebacks
When InferenceClientModel hit a 402 (credits exhausted), the error was a 60-line Python traceback ending with a generic HTTPStatusError. No "Your HF free credits are depleted — add prepaid credits or switch providers." Just a wall of stack frames. New users will stare at this confused.
Summary: The core agent logic is excellent. The operational tooling around it — cost tracking, error UX, runtime safety — is what's missing. If you're building a quick prototype, none of this matters. If you're deploying to production, expect to build these layers yourself. Review collected by and hosted on G2.com.
