Lab Guide — Building Agents
Three exercises building a research agent with mock tools, a ReAct loop, and multi-task testing.
Setup: Open
lab.pdf in Jupyter. Run the setup cell — it loads your API key and defines mock data. The notebook provides MOCK_SEARCH_DB and MOCK_URL_DB dictionaries so the agent runs without real web access.
Exercise 1 — Build Mock Tools
EX 1
Implement 4 tools: search, read_url, save_note, get_notes
Build the tool execution functions and define their tool schemas for Claude's tool-use API.
Tool specifications
| Tool | Parameters | Behaviour |
|---|---|---|
search | query: str | Look up query in MOCK_SEARCH_DB; return matching results or "No results found" |
read_url | url: str | Look up URL in MOCK_URL_DB; return content or "URL not found" |
save_note | key: str, content: str | Store content in a notes dict under key; return confirmation |
get_notes | (none) | Return all saved notes as a formatted string |
Test your implementation
print(execute_tool("search", {"query": "python async"}))
# Should return mock search results
print(execute_tool("save_note", {"key": "test", "content": "hello"}))
# Should confirm note saved
print(execute_tool("get_notes", {}))
# Should show the saved noteCheckpoint: All four tools work correctly with mock data.
Exercise 2 — The ReAct Agent Loop
EX 2
Implement
Implement run_agent(task, tools, max_steps=10)
Build a full ReAct agent loop that:
- Sends the task to Claude with the tool definitions
- Uses a system prompt that instructs Claude to think step by step
- Handles
tool_usestop reason: execute tools, feed results back - Handles
end_turn: extract and return the final text - Enforces
max_stepsguard - Prints each step (thought, action, observation) to the console
System prompt suggestion
SYSTEM = """You are a research agent. You have access to tools for searching,
reading URLs, and saving notes. Follow this process:
1. Think about what information you need
2. Use tools to gather information
3. Save important findings as notes
4. When you have enough information, provide your final answer
Be systematic and thorough."""Test your implementation
result = run_agent(
"Research Python async programming. What are the key concepts? Save your findings.",
TOOLS
)
print("FINAL:", result)Checkpoint: Agent completes the task in multiple steps, using search and save_note tools.
Exercise 3 — Multi-Task Testing
EX 3
Run 3 tasks and analyse step count
Test your agent on three different task types and compare how many steps each requires.
Test tasks
tasks = [
# Research task: requires search + read + save
"Research Python type hints. Find key concepts and save a summary.",
# Comparison task: requires multiple searches + analysis
"Compare Python lists vs tuples. Search for information about both and explain the differences.",
# Planning task: requires gathering info + synthesis
"Create a study plan for learning Python web development. Research the key frameworks and tools needed.",
]What to record
- Number of steps (tool calls) per task
- Which tools were used and in what order
- Whether the agent completed the task or hit max_steps
- Quality of the final answer (subjective 1–5 rating)
Checkpoint: You have a comparison table showing step count, tools used, and quality rating for all 3 tasks.
Tip: If the agent gets stuck in a loop, check your tool result formatting. Make sure tool results are clear and informative — vague results like "OK" do not give the agent enough signal to decide its next step.