	In Browser
	StumbleUpon
	del.icio.us
	Google
	Google Buzz
	reddit
	LinkedIn

	Facebook
	Twitter
	Linkedin
	E-Mail

Agentic AI > CrewAI > CrewAI Agents and Tasks

CrewAI Agents and Tasks

Author: Venkata Sudhakar

CrewAI is a multi-agent framework where you define a team of specialised AI agents, each with a role, goal, and backstory, then assign them tasks to execute in sequence or in parallel. Unlike a single LLM call, a CrewAI crew orchestrates multiple agents where each agent focuses on what it does best. One agent might research a topic, another analyses the findings, and a third writes the final report. Each agent can be given tools, and agents can delegate subtasks to each other.

The three core building blocks are Agent (defines who the agent is - role, goal, backstory, tools), Task (defines what to do - description, expected output, which agent handles it), and Crew (assembles agents and tasks, defines the process - sequential or hierarchical - and kicks off execution). In sequential process, tasks run one after another and each task can use the output of the previous one. In hierarchical process, a manager agent automatically delegates tasks to the best-suited worker agent.

The below example builds a two-agent migration assessment crew: a researcher agent that analyses the migration scope, and a risk analyst agent that produces the final risk report.

# pip install crewai crewai-tools
from crewai import Agent, Task, Crew, Process
from crewai.tools import tool
from langchain_openai import ChatOpenAI

llm = ChatOpenAI(model="gpt-4o-mini", api_key="your-api-key", temperature=0.3)

# Custom tool the agents can use
@tool("Database Schema Analyser")
def analyse_schema(db_description: str) -> str:
    """Analyses a database description and returns table count, stored procedure count and complexity estimate."""
    # In production this would connect to a real database
    return ("Schema analysis: 180 tables, 45 stored procedures, " +
            "12 complex views, estimated complexity: HIGH. " +
            "Oracle-specific features detected: CONNECT BY, DBMS_JOB, and RAW data types.")

# Agent 1: Migration Researcher - scopes the work
researcher = Agent(
    role="Database Migration Researcher",
    goal="Analyse the source database and document all migration challenges",
    backstory=(
        "You are a senior DBA with 15 years of experience in Oracle to PostgreSQL "
        "migrations. You are thorough, technical, and spot problems others miss."
    ),
    tools=[analyse_schema],
    llm=llm,
    verbose=True
)

# Agent 2: Risk Analyst - produces the deliverable
risk_analyst = Agent(
    role="Migration Risk Analyst",
    goal="Produce a concise, actionable risk assessment and timeline recommendation",
    backstory=(
        "You are a project manager specialising in database migration risk. "
        "You translate technical findings into business-readable risk reports."
    ),
    llm=llm,
    verbose=True
)

Defining Tasks and running the Crew,

# Task 1: Research - assigned to the researcher agent
research_task = Task(
    description=(
        "Analyse the following database migration scope and identify all technical challenges: "
        "Oracle 11g ERP system with custom modules migrating to Aurora PostgreSQL. "
        "Use the Database Schema Analyser tool to get schema details. "
        "Document: object counts, Oracle-specific features, estimated conversion effort."
    ),
    expected_output=(
        "A technical findings report listing: table count, stored procedure count, "
        "Oracle-specific features found, and top 3 technical risks."
    ),
    agent=researcher
)

# Task 2: Risk assessment - uses output of Task 1 automatically
risk_task = Task(
    description=(
        "Using the technical findings from the researcher, produce a risk assessment report. "
        "Include: overall risk level (LOW/MEDIUM/HIGH/CRITICAL), timeline recommendation in weeks, "
        "top 3 blockers, and one recommended first action."
    ),
    expected_output=(
        "A structured risk report with risk level, recommended timeline, "
        "blockers list, and immediate next action."
    ),
    agent=risk_analyst
)

# Assemble and run the crew
crew = Crew(
    agents=[researcher, risk_analyst],
    tasks=[research_task, risk_task],
    process=Process.sequential,  # tasks run in order, output flows forward
    verbose=True
)

result = crew.kickoff()
print("\n=== FINAL REPORT ===")
print(result)

It gives the following output,

[Migration Researcher] Using tool: Database Schema Analyser
[Migration Researcher] Schema analysis: 180 tables, 45 stored procedures...
[Migration Researcher] Findings: HIGH complexity Oracle system detected.
  Top risks: (1) CONNECT BY queries need CTEs, (2) DBMS_JOB -> pg_cron,
  (3) RAW data type -> BYTEA conversion

[Risk Analyst] Processing researcher findings...
[Risk Analyst] Producing risk report...

=== FINAL REPORT ===
Risk Level: CRITICAL
Recommended Timeline: 20 weeks
Blockers:
  1. 45 stored procedures require manual PL/pgSQL conversion
  2. CONNECT BY hierarchical queries not supported in PostgreSQL natively
  3. DBMS_JOB scheduler must be replaced with pg_cron
Immediate Next Action: Run AWS Schema Conversion Tool to get auto-conversion
  percentage estimate before committing to timeline.

CrewAI shines for workflows where different steps genuinely need different expertise and context. Use sequential process when tasks have a clear linear dependency - each builds on the previous. Use hierarchical process when you want a manager agent to autonomously decide which worker handles each subtask. Keep each agent focused on one thing: a researcher that researches, an analyst that analyses, a writer that writes. Giving every agent every tool leads to confused, unfocused output.

Send your comments, suggestions or queries regarding this site to [email protected].