Write an agent loop in code with smolagents (sandboxed)

The Stack

Steps

1. 1

   Write the code-agent and structure-check it

   Write agent.py: a CodeAgent with a typed tool, a model via LiteLLM, and executor_type set to a sandbox. CI compiles it and checks the calls. Actually running it needs smolagents, a sandbox, and a key, so that step is fenced.

   cat > agent.py <<'PY'
   from smolagents import CodeAgent, LiteLLMModel, tool
   
   
   @tool
   def get_weather(city: str) -> str:
       "Return the current weather for a city."
       return "sunny in " + city
   
   
   agent = CodeAgent(
       tools=[get_weather],
       model=LiteLLMModel(model_id="anthropic/claude-opus-4.8"),
       executor_type="e2b",
   )
   
   if __name__ == "__main__":
       print(agent.run("What is the weather in Paris?"))
   PY
   python3 - <<'CHECK'
   import py_compile, sys
   src = open("agent.py").read()
   try:
       py_compile.compile("agent.py", doraise=True)
   except py_compile.PyCompileError:
       print("BAD: agent.py does not compile"); sys.exit(1)
   def need(tok, msg):
       if tok not in src:
           print("BAD: " + msg); sys.exit(1)
   need("CodeAgent", "no smolagents CodeAgent")
   need("@tool", "no tool defined")
   need("model=", "no model set")
   sandbox = "e2b" if 'executor_type="e2b"' in src else ("docker" if 'executor_type="docker"' in src else "")
   if not sandbox:
       print("BAD: executor_type is not a sandbox (set e2b or docker)"); sys.exit(1)
   print("config OK: smolagents CodeAgent compiles, defines a tool, names a model, runs in a sandbox (executor_type=" + sandbox + ")")
   CHECK

2. 2

   Run the agent (the model + execution step, not checked by CI)

   pip install smolagents, set your model key and an E2B key (or Docker), and run agent.py. The agent writes Python and the sandbox executes it off your machine. The run is fenced.