Add CodeAgent subagent with Qwen coder model + expose compute tool

New architecture at manager level (4 subagents):
BasicAgent
├── web_research -> WebResearchAgent
├── file_analysis -> FileAnalysisAgent
├── media_analysis -> MediaAgent
├── compute -> CodeAgent (NEW)
└── AnswerNormalizer

CodeAgent wraps python_exec + calculator behind a specialist that
writes Python, runs it, and iterates on errors. Uses Qwen 2.5 Coder
32B by default via GROQ_CODE_MODEL env var (override to
openai/gpt-oss-120b or deepseek-r1-distill-llama-70b if Groq drops
the Qwen model).

tools.py: python_exec + calculator no longer in ALL_TOOLS; wrapped
by compute() which delegates to CodeAgent.

app.py: manager system prompt updated to describe 4 specialists.

Co-Authored-By: Claude Opus 4.7 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_019ryPUj8AtsFopnufGRSc4L

agents.py

@@ -13,6 +13,16 @@ def _make_llm(model: str = "openai/gpt-oss-120b", temperature: float = 0.1) -> C
     )
 
 
+def _make_code_llm() -> ChatGroq:
+    """LLM for CodeAgent. Defaults to Qwen 2.5 Coder 32B on Groq.
+
+    Override with GROQ_CODE_MODEL env var if Groq drops it (e.g. set to
+    ``openai/gpt-oss-120b`` or ``deepseek-r1-distill-llama-70b``).
+    """
+    model = os.getenv("GROQ_CODE_MODEL", "qwen-2.5-coder-32b")
+    return _make_llm(model=model, temperature=0.1)
+
+
 WEB_RESEARCH_PROMPT = """You are a web research specialist.
 
 For each question:
@@ -45,6 +55,22 @@ Workflow:
    verbatim when possible. Do not speculate."""
 
 
+CODE_AGENT_PROMPT = """You are a Python computation specialist.
+
+You answer questions that need arithmetic, list/string manipulation, date math,
+pandas transforms, unit conversions, or any calculation better *computed* than
+reasoned about.
+
+Workflow:
+1. Prefer `calculator` for a single closed-form math expression.
+2. Otherwise write a short Python snippet and run it via `python_exec`.
+   You MUST `print()` any value you want to see back — the tool returns stdout.
+3. If `python_exec` returns an error, read the traceback, fix the code, and
+   retry. Do not give up after one failure; iterate up to 3 times.
+4. Return ONLY the final computed value. No code in the final answer, no
+   explanation, no units unless asked."""
+
+
 MEDIA_ANALYSIS_PROMPT = """You are a media analysis specialist (image, audio, video).
 
 You handle questions that reference a task_id, URL, or file path pointing to
@@ -144,6 +170,21 @@ class MediaAgent:
         return result["messages"][-1].content
 
 
+class CodeAgent:
+    def __init__(self):
+        from tools import calculator, python_exec  # lazy: avoids circular import
+
+        self.graph = create_agent(
+            _make_code_llm(),
+            tools=[python_exec, calculator],
+            system_prompt=CODE_AGENT_PROMPT,
+        )
+
+    def __call__(self, question: str) -> str:
+        result = self.graph.invoke({"messages": [("user", question)]})
+        return result["messages"][-1].content
+
+
 class AnswerNormalizer:
     def __init__(self):
         self.llm = _make_llm(temperature=0.0)

app.py

@@ -17,22 +17,22 @@ from langgraph.prebuilt import ToolNode
 
 MANAGER_SYSTEM_PROMPT = """You are a QA agent for the GAIA benchmark.
 
-You have five tools:
-- web_research(question): factual/current world knowledge (people, dates, events).
+You have four specialist tools:
+- web_research(question): factual / current world knowledge (people, dates,
+  events, statistics).
 - file_analysis(question): question about an attached document
-  (pdf/xlsx/csv/txt/json). Include the task_id or URL in your question.
+  (pdf / xlsx / csv / txt / json). Include the task_id or URL in your question.
 - media_analysis(question): question about an image, audio, or video (also
   YouTube URLs). Include the task_id or URL in your question.
-- python_exec(code): run Python for computation, list ops, date math,
-  string parsing. You must print() the result.
-- calculator(expression): plain math expressions.
+- compute(question): math, arithmetic, list ops, date math, string parsing,
+  unit conversions. State the computation in plain language with the inputs.
 
 Rules:
 1. When a task has an attached file (mentioned by task_id), route to
    file_analysis or media_analysis with the task_id embedded in the question.
 2. Ask each specialist ONE precise question at a time.
-3. Chain tools when needed (e.g. web_research to find a number, then
-   calculator to combine).
+3. Chain tools when needed (e.g. web_research to find numbers, then
+   compute to combine them).
 4. Final answer: give ONLY the value the question asks for. No preamble,
    no explanation."""
 

tools.py

@@ -13,7 +13,7 @@ from langchain_core.tools import tool
 from langchain_groq import ChatGroq
 from tavily import TavilyClient
 
-from agents import FileAnalysisAgent, MediaAgent, WebResearchAgent
+from agents import CodeAgent, FileAnalysisAgent, MediaAgent, WebResearchAgent
 
 DEFAULT_API_URL = "https://agents-course-unit4-scoring.hf.space"
 
@@ -381,6 +381,7 @@ def youtube_transcript(url_or_id: str) -> str:
 _web_agent = WebResearchAgent()
 _file_agent = FileAnalysisAgent()
 _media_agent = MediaAgent()
+_code_agent = CodeAgent()
 
 
 @tool
@@ -431,10 +432,24 @@ def media_analysis(question: str) -> str:
     return _media_agent(question)
 
 
+@tool
+def compute(question: str) -> str:
+    """Delegate math, arithmetic, data manipulation, or string parsing to a code specialist.
+
+    The specialist writes Python, runs it, iterates on errors, and returns the
+    computed value. Use for anything better *computed* than reasoned about:
+    sums / averages / dates / list ops / unit conversions / regex extraction.
+
+    Args:
+        question: What to compute, plainly stated. Include any input data or
+            values needed to perform the computation.
+    """
+    return _code_agent(question)
+
+
 ALL_TOOLS = [
     web_research,
     file_analysis,
     media_analysis,
-    python_exec,
-    calculator,
+    compute,
 ]