Thangavel PrasanthTP

DeepLearning.AI - Multi AI Agent Systems with CrewAI

8 min read

Key resources

Below sections contain the key take aways from each lesson.

Overview

6 key components of multi-agent systems

  1. Role playing

  2. Focus

  3. Tools

  4. Cooperation

  5. Guardrails

  6. Memory

    Interestingly, these components indeed makes for a great employee.

Multi-Agent Collab can be

  • Sequential
  • Parallel
  • Hierarchical
  • Asynchronous

Traditional Software Development vs. AI Software Development

Unlike traditional approach, AI Software Development has

  • Fuzzy inputs
  • Fuzzy transformations
  • Fuzzy outputs
  • Probabilistic nature

Lead Generation: Data Collection & Analysis Example

  • Traditional Automation
    • Website Visitor (e.g., Form) → Lead (e.g., >10 employees) → Sales (Low/Med/High Priority)
  • Agentic Automation
    • Website Visitor → Lead → Crew of AI Agents → Sales, where the
    • Crew of AI Agents could do
      • Research (data collection)
      • Comparison
      • Scoring
      • Talking points

AI Agents

Initial Building Blocks

  • Agents
  • Tasks
  • Crews

Create agents to research and write an article (L2)

This lesson introduces us to the fundamental concepts of multi-agent systems, and get an overview of crewAI framework.

It's been seen that LLMs perform better when they are role playing.

Steps involved in creating a Multi AI Agent

  1. **Create Agents by providing them role, goal, backstory
  2. Create Tasks by providing them description, expected_output, and agent
  3. Create Crew, a top-level organization that oversees the agents and tasks

The Agentic Crew to research and write any article was assembled using 3 Agents and 3 Tasks, which are

  • Content Planner
  • Content Writer
  • Editor For more details, refer this GitHub code

Key elements of AI agents

Agents work better when they are role playing, focused, have appropriate tools, allowed to cooperate, leveraging memory, and having some guardrails.

  1. Role Playing
    • e.g., “Give me an analysis on tesla stock” vs. “you are a FINRA approved financial analyst. give me an analysis on tesla stock”.
      • The later is more effective.
  2. Focus

If we mix things too much (e.g., too many tools, information, context, etc.), our models can loose important information, and hallucinate more.

Do not use one single agent to achieve everything.

  1. Tools

Two different ways to give Agents Tools

  1. Agent Level
  2. Task Level
  1. Cooperation

Ability to cooperate (take feedbacks, delegate tasks) and to bounce ideas of each other produces much better outcome.

  1. Guardrails

Prevent Agent from derailing and budge them to stay on track, thereby making sure to prevent hallucinations and we get reliable and consistent results.

  1. Memory

Memory makes huge, immense difference on our Agent.

3 types of memory offered by CrewAI out of the box:

  1. Short term memory
  2. Long term memory
  3. Entity memory

Multi agent customer support automation (L3)

This lesson covers the six key elements that help make AI Agents perform better.

The Agentic Crew to provide customer support was assembled using 2 Agents and 2 Tasks

  • Agents
    • support_agent with role="Senior Support Representative"
    • support_quality_assurance_agent with role="Support Quality Assurance Specialist" and allow_delegation=True, so that agent can delegate its work to another agent which is better suited to do a particular task.
  • Tool
    • docs_scrape_tool that scrapes a 1 page (URL) of CrewAI documentation.
  • Tasks
    • inquiry_resolution with tools=[docs_scrape_tool] so that the Agent can access this tool on this specific-task.
    • quality_assurance_review
  • Crew
    • crew with memory=True to enable the Memory

Interestingly, the overall token cost of gpt-4o-mini is expensive than gpt-4o, and it even made too many attempts which shows it's limitations.

For more details, refer this GitHub code

Mental framework for agent creation

Think as a Manager, as they are conditioned to think

  • What is the Goal?
  • What is the Process?

What kind of people would I need to hire to get this job done? What should be their roles, goals, backstories?

Those are the agents we want to build.

Example: Use case related to HR

Okay Agent 👎Better Agent 👍
ResearcherHR Research Specialist
WriterSenior Copywriter
Financial AnalystFINRA Approved Analyst

Key elements of agent tools

What makes a great Tool?

  1. Verstaile
  2. Fault-tolerant
  3. Cache

Tools is the connection between AI apps (that have fuzzy inputs) and the external world (that have strongly typed inputs).

CrewAI support Cross-agent caching

For more details, refer [Available CrewAI Tools] (https://docs.crewai.com/en/concepts/tools#available-crewai-tools).

We can also use all the tools that LangChain offers.

Tools for a customer outreach campaign (L4)

This lesson covers more about tools, including custom ones

The Agentic Crew to perform customer outreach campaign (drafting different emails for different target audience) was assembled using 2 Agents, 4 Tools, and 2 Tasks

  • Agents
    • sales_rep_agent with role=Sales Representative
    • lead_sales_rep_agent with role=Lead Sales Representative
  • Tool
    • directory_read_tool that list files in directory
    • file_read_tool that reads file contents
    • search_tool (from SerperDevTool) to search the internet with Serper
    • sentiment_analysis_tool (custom tool) that’s hardcoded output
  • Tasks
    • lead_profiling_task with tools=[directory_read_tool, file_read_tool, search_tool]
    • personalized_outreach_task with tools=[sentiment_analysis_tool, search_tool]
  • Crew
    • crew with memory=True to enable the Memory

For more details, refer this GitHub code

Key elements of well defined tasks

What tasks and processes do I expect the individuals on my team to do?

Those are the tasks we want to create.

Two key elements of well defined tasks

  1. Clear description of the tasks
  2. Set a clear and concise expectation

Other (hyper)parameters or task attributes

  • Set context
    • By using context, we can specify other tasks whose outputs will be used as context for this task.
  • Set a callback
    • By using callback, we can specify the Function/object to be executed after task completion.
  • Override Agent tools with specific task tools
    • By using tools, we can specify the tools/resources the agent is limited to use for this task.
  • Force human input before end of task
    • By setting human_input=True, the task will ask for human feedback (whether you like the results or not) before finalising it.
  • Execute asynchronously
  • Output as Pydantic
    • By using output_pydantic
  • Output as JSON
    • By using output_json, you can specify the structure of the output you want.
  • Output as file
    • By using output_file, you can get your output in a file.
  • Run in parallel
    • By using async_execution=True, we can specify the task should be executed asynchronously. Defaults to False.

For more details, refer crew ai documentation > tasks

Automate Event Planning (L5)

This lesson covers about tasks, and different parameters for it.

The Agentic Crew to perform automated event planning was assembled using 3 Agents, 2 Tools, 3 Tasks

  • Agents
    • venue_coordinator with tools=[search_tool, scrape_tool]
    • logistics_manager with tools=[search_tool, scrape_tool]
    • marketing_communications_agentwith tools=[search_tool, scrape_tool]

Interestingly, tools are passed directly into Agents unlike previous examples, where they are passed only to the tools

  • Class VenueDetails using Pydantic BaseModel
  • Tasks
    • venue_task with human_input=True, output_json=VenueDetails, and output_file="venue_details.json"
    • logistics_task with human_input=True
    • marketing_task with async_execution=True and output_file="marketing_report.md"
  • Crew

The human_input=True turned out to be useful for providing feedback on the Final Result and Agent's actions.

For more details, refer this GitHub code

Multi agent collaboration for financial analysis (L6)

This lesson covers on ways to make agents collaborate with each other.

The multi Agentic Crew to perform financial analsyis was assembled using 4 Agents, 2 Tools, and 3 Tasks

  • Agents
    • data_analyst_agent
    • trading_strategy_agent
    • execution_agent
    • risk_management_agent Note: All the Agents were given allow_delegation=True, and tools=[scrape_tool, search_tool]
  • Tasks
    • data_analysis_task
    • strategy_development_task
    • execution_planning_task
    • risk_assessment_task Note: All the Tasks were configured with output_file to output a markdown file.
  • Crew
    • crew with process=Process.hierarchical

Why hierarchical process instead of sequential process?

For a financial trading crew with multiple specialized agents, the hierarchical process is likely more appropriate as it allows for better coordination between the data analyst, strategy developer, execution agent, and risk manager, rather than forcing them to work in strict sequence.

For more details, refer this GitHub code

Build a crew to tailor job applications (L7)

This lesson covers building our first multi-agent system

The multi Agentic Crew to perform tailored job applications using 4 Agents, 3

  • Agents
    • researcher with role="Tech Job Researcher"
    • profiler with role="Personal Profiler for Engineers"
    • resume_strategist with role="Resume Strategist for Engineers"
    • interview_preparer with role="Engineering Interview Preparer" Note: All the Agents were given tools=[scrape_tool, search_tool, read_resume, semantic_search_resume], except researcher which was only given tools=[scrape_tool, search_tool]
  • Tasks
    • research_task with output_file as markdown and async_execution=True to extract job requirements
    • profile_task with output_file as markdown and async_execution=True to compile comprehensive profile
    • resume_strategy_task with output_file as markdown and context=[research_task, profile_task] to align resume with job requirements
    • interview_preparation_task with output_file as markdown and context=[research_task, profile_task, resume_strategy_task] to develop interview materials

How does context work in Tasks?

  • You can pass a list of tasks as context to a task.
  • The task then takes into account the output of those tasks in its execution.
  • The task will not run until it has the output(s) from those tasks.

Next steps with AI Agent Systems

Check out these resources

To do or clarify

  • Multi agent customer support automation (L3)
    • When we set Crew(..., memory=True), does it enable short-term memory or long-term memory? How about entity memory?
    • Citation is not provided in the demo output. Is it a glitch or something else?
      • Seems this is resolved once migrated to latest GPT 4 variant models
    • What’s considered as guardrail in this project?
      • Is it the expected_output defined in the Task or backstory provided in the Agent or something else?
  • Tools for a customer outreach campgain (L4)
    • Does this tool has caching with a timeout so that we don’t have to pay again for the same call for same query in a short span?
    • What’s cached_promot_tokens in result.token-usage
      • For e.g., Token Usage: total_tokens=67180 prompt_tokens=65040 cached_prompt_tokens=51200 completion_tokens=2140 successful_requests=15 for gpt-4o
  • Automate Event Planning (L5)
    • When async_execution=True was enabled for two tasks, it’s throwing an error. How to fix it?

Graph View

Backlinks