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

	Facebook
	Twitter
	Linkedin
	E-Mail

Agentic AI > CrewAI > Multi-Agent Systems with CrewAI

Multi-Agent Systems with CrewAI

Author: Venkata Sudhakar

CrewAI is an open-source framework for building multi-agent AI systems where multiple specialised AI agents collaborate to complete complex tasks. The key insight is that a single LLM agent trying to do everything often produces mediocre results across all subtasks, whereas multiple specialised agents - each with a clear role, goal, and set of tools - produce much better results by division of labour. CrewAI models this as a "crew" of agents working together, similar to how a human team operates.

The three core CrewAI concepts are: Agent (an LLM with a specific role, goal, backstory, and set of tools), Task (a specific piece of work assigned to an agent, with a description of the expected output), and Crew (the team of agents and list of tasks, with a process that defines how they collaborate). The two main process types are sequential (tasks run one after another, output of each feeding into the next) and hierarchical (a manager agent delegates tasks to worker agents and synthesises the results). CrewAI is built on top of LangChain, so all LangChain tools work with CrewAI agents.

The below example shows a CrewAI crew for data migration planning, with three specialised agents: a migration analyst, a risk assessor, and a report writer working sequentially.

# pip install crewai crewai-tools
from crewai import Agent, Task, Crew, Process
from crewai_tools import SerperDevTool
from langchain_openai import ChatOpenAI

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

# Define specialised agents with distinct roles
migration_analyst = Agent(
    role="Senior Data Migration Analyst",
    goal="Analyse the scope and technical requirements of database migrations",
    backstory=("You are an expert in database migrations with 10 years of experience "
               "migrating Oracle, MySQL and SQL Server databases to the cloud. "
               "You specialise in assessing data volumes, dependencies and migration strategies."),
    verbose=True,
    llm=llm
)

risk_assessor = Agent(
    role="Migration Risk Assessor",
    goal="Identify and quantify risks in database migration projects",
    backstory=("You are a risk management expert for enterprise IT projects. "
               "You have seen hundreds of migrations fail due to data quality issues, "
               "schema incompatibilities, and underestimated downtime. "
               "You identify risks others overlook."),
    verbose=True,
    llm=llm
)

report_writer = Agent(
    role="Technical Report Writer",
    goal="Produce clear, executive-ready migration planning reports",
    backstory=("You translate complex technical migration analyses into "
               "structured reports that both engineers and business stakeholders can act on. "
               "You are concise, precise, and always recommend a clear next step."),
    verbose=True,
    llm=llm
)

# Define tasks assigned to each agent
analysis_task = Task(
    description=("Analyse a migration from on-premises MySQL 5.7 with 50 million rows "
                 "across 120 tables to Cloud Spanner on GCP. "
                 "Identify the key technical challenges: data type incompatibilities, "
                 "stored procedures, and triggers that do not exist in Spanner."),
    expected_output="A structured technical analysis with sections for scope, challenges and recommended approach.",
    agent=migration_analyst
)

risk_task = Task(
    description=("Based on the migration analysis, identify the top 5 risks "
                 "for the MySQL to Cloud Spanner migration. "
                 "For each risk state: probability (high/medium/low), "
                 "impact (high/medium/low), and mitigation strategy."),
    expected_output="A risk register with 5 risks, each with probability, impact and mitigation.",
    agent=risk_assessor,
    context=[analysis_task]  # This task receives the output of analysis_task
)

report_task = Task(
    description=("Produce a one-page executive migration planning report "
                 "that combines the technical analysis and risk register. "
                 "Include: executive summary, recommendation, timeline estimate, "
                 "and the 3 most critical actions the team must take immediately."),
    expected_output="A formatted one-page executive report ready to present to stakeholders.",
    agent=report_writer,
    context=[analysis_task, risk_task]
)

# Assemble and run the crew
crew = Crew(
    agents=[migration_analyst, risk_assessor, report_writer],
    tasks=[analysis_task, risk_task, report_task],
    process=Process.sequential,
    verbose=True
)

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

It gives the following output (abbreviated),

[Senior Data Migration Analyst]: Analysing MySQL 5.7 to Cloud Spanner migration...
Key findings:
- 50M rows across 120 tables: estimated 8-12 hours initial load
- 23 stored procedures must be rewritten (Spanner has no stored procs)
- AUTO_INCREMENT not supported in Spanner (use UUID or sequence emulation)
- JSON columns require schema change to STRING or ARRAY types
- Recommended: phased CDC-based migration over 12 weeks

[Migration Risk Assessor]: Identifying top 5 risks...
Risk 1: Stored procedure rewrite failures - PROBABILITY: High, IMPACT: High
  Mitigation: Audit all 23 procs in week 1, assign dedicated dev for rewrites
Risk 2: Schema incompatibility data corruption - PROBABILITY: Medium, IMPACT: High
  Mitigation: Run parallel validation for 2 weeks before cutover
...

=== FINAL REPORT ===
EXECUTIVE SUMMARY
The proposed MySQL 5.7 to Cloud Spanner migration involves 50M rows and presents
high complexity due to 23 stored procedures and schema incompatibilities.
RECOMMENDATION: Proceed with phased 12-week migration using Datastream CDC.
TIMELINE: 12 weeks. IMMEDIATE ACTIONS:
1. Complete stored procedure audit by end of Week 1.
2. Provision Cloud Spanner instance and run schema compatibility validator.
3. Set up Datastream CDC pilot on 3 non-critical tables.

CrewAI vs LangGraph - when to choose each:

Choose CrewAI when you want a high-level, role-based abstraction for multi-agent collaboration. CrewAI is easier to set up, reads naturally like a team structure, and is ideal for content generation, research, and analysis workflows where different "experts" contribute their speciality. Choose LangGraph when you need fine-grained control over state, complex conditional routing, cycles, or human-in-the-loop interrupts. LangGraph is lower level but more powerful for building production agents where you need to debug exactly what is happening at each step.

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