smolagents Deep Dive: HuggingFace's 1,000-Line Agent Loop in 2026

TL;DR — smolagents is the most readable agent framework in the ecosystem. The core loop is under 1,000 lines, the CodeAgent writes Python instead of JSON tool calls, and on GAIA it scores 44.2% versus GPT-4 Turbo's 7%. If you want to understand what your agent is doing, start here.

Why smolagents

flowchart LR
  Repo[GitHub repo] --> CI[GitHub Actions]
  CI --> Eval[Agent eval suite · PromptFoo]
  Eval -->|pass| Deploy[Deploy]
  Eval -->|fail| Block[Block PR]
  Deploy --> Prod[Production agent]
  Prod --> Trace[(LangSmith trace)]
  Trace --> Eval

HuggingFace shipped smolagents in late 2024 as a deliberate counter-reaction to the megaframework era. The pitch: an agent doesn't need a planner, a router, a memory module, and a 50-class type hierarchy. It needs a loop that picks a tool and runs it. agents.py weighs in at under a thousand lines. The library went from 3k stars in early 2025 to 26k+ by April 2026.

The core loop

A smolagent does exactly this:

1. Format the system prompt with the available tools.
2. Send the conversation to the LLM.
3. Parse the response. If it's a final answer, return. If it's a tool call (or Python code), execute it.
4. Append the result and loop until max_steps or a final answer.

That's it. The two flagship agent classes are:

• CodeAgent — the model emits Python code directly. The framework executes it in a sandbox and returns the result. Research shows code agents use 30% fewer steps than JSON tool-calling agents and score higher on hard benchmarks.
• ToolCallingAgent — the classic JSON tool-call loop, for models that don't write Python well or for environments where you can't sandbox.

Sandboxed execution is non-negotiable

Letting a model emit arbitrary Python and exec()-ing it on your laptop is a textbook prompt-injection RCE. smolagents ships first-class executors for E2B, Modal, Blaxel, Docker, and the Pyodide + Deno WebAssembly sandbox. You opt in with one parameter:

from smolagents import CodeAgent, HfApiModel

agent = CodeAgent(
    tools=[search_tool, fetch_tool],
    model=HfApiModel("Qwen/Qwen2.5-72B-Instruct"),
    executor_type="e2b",   # or "modal", "docker", "wasm"
    additional_authorized_imports=["pandas", "numpy"],
)

For CallSphere internal use we lean on Modal for batch agent workloads — it gives us per-task containers with GPU optionality and no idle cost.

Where smolagents shines (and where it doesn't)

Pick smolagents when:

• You want to read every line of the loop and understand what's happening.
• Your agent's job is code, data, or tool orchestration — anything where Python is the natural action language.
• You need a portable agent that runs locally with Ollama, on HF Inference, with OpenAI, with Anthropic, or with LiteLLM, with no code changes.
• You're teaching a team how agents work without burying them in abstractions.

Skip smolagents when:

• You need persistent agent memory across sessions (use Letta or mem0 for that).
• You need graph-based topology with branching/joining (use LangGraph).
• You need a production observability stack out of the box (smolagents has hooks but no batteries — pair with Arize Phoenix or Langfuse).

How CallSphere uses it

CallSphere's voice surface is OpenAI Realtime + the OpenAI Agents SDK orchestrator (37 specialist agents, 90+ tools, 115+ DB tables across 6 verticals). For non-voice batch workloads — the GTM scrapers, the SEO content generators, the affiliate-fraud detectors — we run smolagents CodeAgents on Modal because the loop is auditable and the cost per task is pennies. The GAIA-style "open-ended research with tools" pattern is exactly what these batch jobs need.

Explore a live demo and compare current plans to find the right fit for your business.

Build steps — your first production smolagent

1. pip install smolagents[e2b] and set E2B_API_KEY.
2. Define tools as plain Python functions decorated with @tool — type hints become the JSON schema automatically.
3. Pick a model: HfApiModel, LiteLLMModel, OpenAIServerModel, or TransformersModel for local.
4. Set max_steps low (5-8) for online workloads, higher (20+) for offline research.
5. Wire telemetry via the OpenInference instrumentor so traces flow to Phoenix, Langfuse, or LangSmith.
6. Add a guardrail tool that the agent must call before final answer (e.g., validate_output(...)).
7. Deploy to Modal or E2B; set a per-task budget cap.

Code: a CodeAgent with sandboxed tools

from smolagents import CodeAgent, LiteLLMModel, tool

@tool
def fetch_competitor_pricing(domain: str) -> str:
    """Fetch a competitor's public pricing page and return the raw HTML."""
    import httpx
    return httpx.get(f"https://{domain}/pricing", timeout=10).text

@tool
def parse_table(html: str, css_selector: str) -> list[dict]:
    """Parse an HTML table into a list of dicts using the given CSS selector."""
    from bs4 import BeautifulSoup
    soup = BeautifulSoup(html, "html.parser")
    return [{c.get("data-key", c.text): c.text for c in row.find_all("td")}
            for row in soup.select(css_selector)]

agent = CodeAgent(
    tools=[fetch_competitor_pricing, parse_table],
    model=LiteLLMModel("claude-sonnet-4"),
    executor_type="modal",
    additional_authorized_imports=["pandas", "json"],
    max_steps=8,
)
result = agent.run("Compare CallSphere's $19/$49/$99/$149 tiers to a competitor's published pricing.")

The CodeAgent's plan, in our traces, looks something like: fetch the page, parse the pricing table, build a pandas DataFrame, compute deltas, return a summary. The model never emits JSON tool calls — it writes Python that orchestrates the tools. That's the 30% step-count win.

Production gotchas we hit

Three things bit us when we first shipped smolagents:

1. additional_authorized_imports is a kill-switch. If you don't list a stdlib module the agent needs (e.g., json, re, datetime), every code execution fails. Start permissive in dev, tighten in prod.
2. Token costs creep on tool descriptions. The CodeAgent prompt re-emits every tool's signature each turn. We trim docstrings to one line in production prompts.
3. max_steps is critical. Without it a confused agent will burn dollars in a loop. We default to 5-8 for online jobs and gate higher values behind a feature flag.

FAQ

Is the CodeAgent really safe? Only if you sandbox. Never run a CodeAgent's code in your application process. The framework will warn you, but it's your responsibility to set executor_type.

Can I use smolagents with MCP servers? Yes — there are MCP client wrappers in the community. As of April 2026 the official smolagents repo has examples that mount MCP tools as @tool functions.

How does it compare to Pydantic AI? Pydantic AI is type-first; smolagents is loop-first. If you want strict structured outputs from every step, Pydantic AI. If you want a transparent ReAct loop, smolagents.

How do I evaluate it? Use Promptfoo or Phoenix Evals; smolagents emits standard OpenInference spans.

Does smolagents support multi-agent topologies? Yes — agents can spawn and call other agents as tools. The pattern is simpler than CrewAI's role-playing model but plenty for most production cases.

What model picks should I default to? GPT-5 or Claude Sonnet 4 for production CodeAgents. For local, Qwen2.5-72B-Instruct is the sweet spot of cost and Python-writing quality.

How do I plug in OpenTelemetry? Use openinference-instrumentation-smolagents. One line, traces flow to Phoenix, Langfuse, or your OTel backend.

Sources

• smolagents on GitHub
• Introducing smolagents (HF blog)
• Secure code execution docs
• smolagents 26k+ stars deep-dive