Agent Foundry
All Problems

#27. Composable Tool Pipeline

HardTool CallingOrchestration

The Problem

Your data analyst agent has a single monolithic tool called process_data that does everything: fetching raw data, cleaning it, running analysis, and formatting a report. This makes it impossible to reuse individual steps — if a user only wants to fetch data without analysis, or run a different type of analysis on already-fetched data, the entire pipeline runs anyway. Your job is to decompose the monolith into small, single-purpose tools that the agent can compose into different pipelines depending on the query.

Examples

Example 1

User input: Get financial data for Acme Corp and analyze their growth

Current (bad) output: The monolithic tool runs all four stages regardless. If the user later asks for a financial analysis (not growth) on the same data, the tool re-fetches and re-cleans everything.

Expected (good) output: The agent calls fetch_data("Acme Corp") → clean_data(raw_data) → analyze_data(cleaned_data, "growth") → format_report(analysis). Each step is separate, and the intermediate results could be reused.

Example 2

User input: I already have this data: "revenue: $10M, costs: $6M". Just run a financial analysis.

Current (bad) output: The monolithic tool ignores the user's data and fetches fresh data from scratch, running the full pipeline.

Expected (good) output: The agent skips fetch_data and clean_data, directly calls analyze_data(user_data, "financial") → format_report(analysis). Only the needed steps are executed.

Example 3

User input: Fetch data for TechCorp but don't analyze it yet, just show me the raw numbers.

Expected (good) output: The agent calls only fetch_data("TechCorp") and clean_data(raw), then presents the cleaned data without running analysis or formatting a full report.

Your Task

  • Break the monolithic process_data tool into at least 4 small tools (e.g., fetch, clean, analyze, format).
  • Each tool should accept the output of the previous step and be independently callable.
  • The agent should dynamically decide which tools to compose based on the user's query.
  • Ensure the decomposed pipeline produces correct results equivalent to the monolithic version.

Evaluation

Submissions are checked for the following:

  • Monolith is decomposed: The single tool is broken into at least 4 independent, single-purpose tools.
  • Tools are composable: The agent can chain tools in different orders depending on the query.
  • Each tool works independently: Any single tool can be called on its own without the full pipeline.
  • Output is still correct: The decomposed pipeline produces equivalent analysis quality.

Constraints

  • The monolithic tool must be broken into at least 4 small, single-purpose tools
  • Each small tool must be independently usable by the agent
  • The agent must be able to compose tools into different pipelines depending on the query
  • No tool should duplicate logic from another
Starter Code
from langchain_openai import ChatOpenAI
from langchain_core.tools import tool
from langgraph.prebuilt import create_react_agent

llm = ChatOpenAI(model="gpt-4o-mini")

# BUG: This monolithic tool does everything — fetch, clean, analyze, and format.
# It's impossible to reuse individual steps or compose different pipelines.
# TODO: Break into small composable tools and build a dynamic LangGraph pipeline.

@tool
def process_data(source: str, analysis_type: str) -> str:
    """Fetch data from source, clean it, analyze it, and format a report."""
    raw_data = f"Raw data from {source}: [revenue: $10M, costs: $6M, employees: 500, growth: 15%, region: US, founded: 2010]"
    cleaned = raw_data.replace("[", "").replace("]", "").strip()
    if analysis_type == "financial":
        analysis = "Profit margin: 40%, Revenue per employee: $20K, Healthy financials."
    elif analysis_type == "growth":
        analysis = "15% YoY growth, above industry average of 10%. Strong trajectory."
    else:
        analysis = "General overview: mid-size company with solid fundamentals."
    report = f"=== Report for {source} ===\nData: {cleaned}\nAnalysis: {analysis}\n=================="
    return report

agent = create_react_agent(llm, [process_data])

result = agent.invoke({"messages": [("human", "Get financial data for Acme Corp and analyze their growth")]})
print(result["messages"][-1].content)
Open in Google Colab
Evaluation Criteria0/4