Equipping Agents with Tools

Built-in Tools

from crewai_tools import SerperDevTool, ScrapeWebsiteTool, FileReadTool

search_tool = SerperDevTool()
scrape_tool = ScrapeWebsiteTool()
file_tool = FileReadTool()

agent = Agent(
    role="Researcher",
    goal="Find accurate information",
    tools=[search_tool, scrape_tool, file_tool],
)

Custom Tools

from crewai.tools import tool

@tool("Database Query")
def query_database(sql: str) -> str:
    """Execute a SQL query against the analytics database.
    Use this to retrieve data about users, orders, or products."""
    return execute_query(sql)

agent = Agent(
    role="Data Analyst",
    tools=[query_database],
)

Tool Design Principles

• Clear descriptions — agents use the docstring to decide when to use a tool
• Specific inputs — define parameter types and constraints
• Useful outputs — return structured, actionable data
• Error messages — return helpful errors, not stack traces

Agent Configuration

Fine-tune agent behavior:

• allow_delegation: Let agents pass tasks to other agents
• max_iter: Maximum reasoning iterations before stopping
• max_rpm: Rate limit for API calls
• memory: Enable conversation memory across tasks

Common Pitfall: Missing Tool Docstrings

Agents use the tool's docstring to decide when and how to use it. A tool with no docstring or a vague docstring like """Use this tool.""" will be called incorrectly or not at all. Write docstrings that answer three questions: What does it do? What inputs does it expect? What does it return?

@tool("Search Company Database")
def search_companies(name: str) -> str:
    """
    Search the internal company database for organizations matching the given name.
    Input: A company name or partial name (e.g. 'Google', 'Open').
    Returns: A JSON list of matching companies with fields: id, name, industry, revenue_usd.
    Use this when the user mentions a specific company and you need its financial details.
    """
    return query_db(f"SELECT * FROM companies WHERE name ILIKE '%{name}%'")

Best Practice: Log Task Outputs for Debugging

Crew task outputs are available after crew.kickoff() completes via the output attribute of each task:

result = crew.kickoff()
for task in crew.tasks:
    print(f"--- {task.description[:50]} ---")
    print(task.output.raw)
    print()

This makes it easy to identify which agent produced a poor result without re-running the entire crew.

Named Product Reference: Groq for Fast CrewAI Inference

CrewAI supports any LangChain-compatible LLM. Groq's inference API (groq.com) runs open-weight models like Llama 3.1 at speeds of 300–800 tokens/second — 10x faster than standard GPU inference. For CrewAI workflows where multiple agents take many turns, Groq's speed dramatically reduces wall-clock time. Use it by setting llm="groq/llama-3.1-70b-versatile" in agent configuration.