Switch from CodeAgent to ToolCallingAgent

15% run showed gpt-oss-120b frequently breaking CodeAgent's <code> tag parser, burning steps on parse failures instead of progress. ToolCallingAgent uses native structured tool calls instead, eliminating this failure mode. PythonInterpreterTool added explicitly to retain compute capability.

app.py

@@ -10,8 +10,9 @@ import spaces
 
 from smolagents import (
     ActionStep,
-    CodeAgent,
     LiteLLMModel,
+    PythonInterpreterTool,
+    ToolCallingAgent,
     WebSearchTool,
     VisitWebpageTool,
     WikipediaSearchTool,
@@ -56,7 +57,7 @@ class TokenPacer:
         self.tokens_per_minute_budget = tokens_per_minute_budget
         self._usage_window: deque[tuple[float, int]] = deque()
 
-    def __call__(self, memory_step: ActionStep, agent: CodeAgent) -> None:
+    def __call__(self, memory_step: ActionStep, agent: ToolCallingAgent) -> None:
         now = time.monotonic()
         usage = getattr(memory_step, "token_usage", None)
         tokens = (usage.input_tokens + usage.output_tokens) if usage else 0
@@ -75,7 +76,7 @@ class TokenPacer:
 class MemoryTrimmer:
     """
     Step_callback that collapses old tool outputs in the agent's memory.
-    CodeAgent re-sends the *entire* step history on every call, so input
+    The agent re-sends the *entire* step history on every call, so input
     tokens grow every single step (2k -> 5k -> 8k -> ... -> 20k+ by step 6),
     which both wrecks the TPM budget and slows every later step far more
     than it needs to. Keeps the most recent `keep_recent` steps' tool
@@ -87,7 +88,7 @@ class MemoryTrimmer:
         self.keep_recent = keep_recent
         self.max_old_observation_chars = max_old_observation_chars
 
-    def __call__(self, memory_step: ActionStep, agent: CodeAgent) -> None:
+    def __call__(self, memory_step: ActionStep, agent: ToolCallingAgent) -> None:
         action_steps = [step for step in agent.memory.steps if isinstance(step, ActionStep)]
         for step in action_steps[:-self.keep_recent]:
             observations = getattr(step, "observations", None)
@@ -116,19 +117,25 @@ class BasicAgent:
             requests_per_minute=float(os.getenv("RATE_LIMIT_RPM", "4.5")),
             retry=False,
         )
-        self.agent = CodeAgent(
+        # ToolCallingAgent (not CodeAgent): gpt-oss-120b kept mixing reasoning
+        # text into its code instead of cleanly wrapping it in <code> tags,
+        # so CodeAgent's regex-based code-block parser failed on a large
+        # fraction of steps - wasted steps that burned tokens/rate-limit
+        # budget without making progress, and led to hallucinated final
+        # answers. ToolCallingAgent uses the provider's native structured
+        # tool-calling instead of parsing free-form text, which sidesteps
+        # this failure mode entirely. PythonInterpreterTool replaces the
+        # code-execution capability CodeAgent had built in.
+        self.agent = ToolCallingAgent(
             model=self.model,
             tools=[
                 WebSearchTool(),
                 VisitWebpageTool(max_output_length=3000),  # default 40000 chars blows the TPM budget in one call
                 WikipediaSearchTool(content_type="summary"),  # "text" (default) returns the full article
-            ],
-            # add_base_tools=True would add a duplicate DuckDuckGo search tool -
-            # every tool's schema is baked into the system prompt on every call,
-            # keeping that overhead small helps stay well under the TPM budget.
-            additional_authorized_imports=[
-                "pandas", "numpy", "math", "re", "json", "itertools",
-                "collections", "statistics", "datetime", "io", "openpyxl", "PIL",
+                PythonInterpreterTool(authorized_imports=[
+                    "pandas", "numpy", "math", "re", "json", "itertools",
+                    "collections", "statistics", "datetime", "io", "openpyxl", "PIL",
+                ]),
             ],
             max_steps=7,  # keep runs short: growing history otherwise blows the TPM budget by step ~6
             step_callbacks=[