Switch back to Claude via LiteLLM, add step rate limiter

Free HF Inference Providers OAuth access was 403ing; switch back to LiteLLMModel/ANTHROPIC_API_KEY, add a step_callback rate limiter to avoid provider throttling across a full eval run.

app.py

@@ -1,5 +1,6 @@
 import os
 import tempfile
+import time
 import gradio as gr
 import requests
 import inspect
@@ -7,8 +8,9 @@ import pandas as pd
 import spaces
 
 from smolagents import (
+    ActionStep,
     CodeAgent,
-    InferenceClientModel,
+    LiteLLMModel,
     WebSearchTool,
     VisitWebpageTool,
     WikipediaSearchTool,
@@ -22,7 +24,7 @@ DEFAULT_API_URL = "https://agents-course-unit4-scoring.hf.space"
 @spaces.GPU
 def _zerogpu_startup_check():
     # This Space runs on ZeroGPU hardware but the agent below only makes
-    # network calls (HF Inference API, web search) and never touches CUDA.
+    # network calls (Claude API, web search) and never touches CUDA.
     # ZeroGPU requires at least one @spaces.GPU function to be declared,
     # so this no-op satisfies that check without spending any GPU quota
     # (it is never actually invoked).
@@ -37,20 +39,36 @@ If you are asked for a string, don't use articles or abbreviations (e.g. for cit
 If you are asked for a comma separated list, apply the above rules to each element depending on whether it's a number or a string.
 """
 
+class RateLimiter:
+    """
+    Counts agent LLM calls (one per step) and sleeps for a fixed duration
+    once a threshold is reached, then resets the counter. Used as a
+    step_callback to stay under the model provider's rate limits across a
+    full run of ~20 GAIA questions, each potentially taking several steps.
+    """
+    def __init__(self, calls_per_wait: int = 15, seconds_to_wait: int = 30):
+        self.calls_per_wait = calls_per_wait
+        self.seconds_to_wait = seconds_to_wait
+        self._call_count = 0
+
+    def __call__(self, memory_step: ActionStep, agent: CodeAgent) -> None:
+        self._call_count += 1
+        if self._call_count >= self.calls_per_wait:
+            print(f"Rate limiter: {self.calls_per_wait} calls reached, sleeping {self.seconds_to_wait}s")
+            time.sleep(self.seconds_to_wait)
+            self._call_count = 0
+
+
 # --- Basic Agent Definition ---
 # ----- THIS IS WERE YOU CAN BUILD WHAT YOU WANT ------
 class BasicAgent:
-    def __init__(self, hf_token: str | None = None):
-        # Uses the logged-in user's own HF Inference Providers quota (via the
-        # "Sign in with Hugging Face" OAuth token, scope: inference-api).
-        # Falls back to the HF_TOKEN env var (e.g. for local testing) if no
-        # OAuth token is available.
-        model_kwargs = {"token": hf_token or os.getenv("HF_TOKEN"), "temperature": 0}
-        model_id = os.getenv("AGENT_MODEL_ID")
-        if model_id:
-            model_kwargs["model_id"] = model_id
-
-        self.model = InferenceClientModel(**model_kwargs)
+    def __init__(self):
+        model_id = os.getenv("AGENT_MODEL_ID", "anthropic/claude-sonnet-5")
+        api_key = os.getenv("ANTHROPIC_API_KEY")
+        if not api_key:
+            print("Warning: ANTHROPIC_API_KEY is not set - the agent will fail to call the model.")
+
+        self.model = LiteLLMModel(model_id=model_id, api_key=api_key, temperature=0)
         self.agent = CodeAgent(
             model=self.model,
             tools=[WebSearchTool(), VisitWebpageTool(), WikipediaSearchTool()],
@@ -60,6 +78,7 @@ class BasicAgent:
                 "collections", "statistics", "datetime", "io", "openpyxl", "PIL",
             ],
             max_steps=12,
+            step_callbacks=[RateLimiter()],
         )
         print("BasicAgent initialized.")
 
@@ -80,7 +99,7 @@ class BasicAgent:
         print(f"Agent returning answer: {answer}")
         return answer
 
-def run_and_submit_all( profile: gr.OAuthProfile | None, oauth_token: gr.OAuthToken | None):
+def run_and_submit_all( profile: gr.OAuthProfile | None):
     """
     Fetches all questions, runs the BasicAgent on them, submits all answers,
     and displays the results.
@@ -101,7 +120,7 @@ def run_and_submit_all( profile: gr.OAuthProfile | None, oauth_token: gr.OAuthTo
 
     # 1. Instantiate Agent ( modify this part to create your agent)
     try:
-        agent = BasicAgent(hf_token=oauth_token.token if oauth_token else None)
+        agent = BasicAgent()
     except Exception as e:
         print(f"Error instantiating agent: {e}")
         return f"Error initializing agent: {e}", None
@@ -232,7 +251,7 @@ with gr.Blocks() as demo:
         Once clicking on the "submit button, it can take quite some time ( this is the time for the agent to go through all the questions).
         This space provides a basic setup and is intentionally sub-optimal to encourage you to develop your own, more robust solution. For instance for the delay process of the submit button, a solution could be to cache the answers and submit in a seperate action or even to answer the questions in async.
 
-        **Setup:** This agent calls Hugging Face's Inference Providers via `smolagents`, using your own account's quota through the login below (no separate API key needed). No hardware/model access beyond your normal HF account is required.
+        **Setup:** This agent calls Anthropic's Claude via `smolagents`. Set the `ANTHROPIC_API_KEY` secret in this Space's settings before running.
         """
     )