DeepLearning.AI - Practical Multi AI Agents and Advanced Use Cases with crewAI

Key resources

• Course link
  • Forum link
• My Github Repo link

Below sections contain the key take aways from each lesson.

---

Overview of Multi AI-Agent Systems

Use cases to be explored in this course

• Automated Project Planning
• Project Monitoring
• Lead Qualification and Scoring
• Analysis and Report on Support Data
• Custom Content Creation at Scale

Most of the use cases usually fo ground some form of research and analysis, summarization and reporting

Main building blocks for building multi AI-agent systems

We can orchestrate agents in different ways

We can build Agents and Crews even with a YAML file

---

Automated Project: Planning, Estimation and Allocation (L2)

Unlike the previous course, in this course, the Agent (backstory, role, and goal) and Task (description, and expected_output) are provided as .yaml files.

The Agentic Crew to perform Project Planning, Estimation and Resource Allocation was assembled using 3 Agents, and 3 Tasks.

• Agents
  • project_planning_agent with role as The Ultimate Project Planner
  • estimation_agent with role as Expert Estimation Analyst
  • resource_allocation_agent with role as Resource Allocation Strategist
• Class TaskEstimate, MileStone, and ProjectPlan using pydantic BaseModel
• Tasks
  • task_breakdown
  • time_resource_estimation
  • resource_allocation with output_pydantic=ProjectPlan
• Crew

Since the output of resource_allocation tasks is configured to be pydantic, it can easily be converted to json and as pandas DataFrame subsequently for downstream processing.

---

Internal and External Agents

Internal tools could be

• RAG search
• SQL query
• Trigger side effect

External tools could be

• Search the internet
• Check calendar
• Reply Email

---

Building Project Progress Report (L4)

The Agentic Crew to create a project progress report was assembled using 2 Agents, and 3 Tasks, and 2 custom built Tools.

• Tools
  • BoardDataFetcherTool and CardDataFetcherTool using crewai BaseTool
• Agents
  • data_collection_agent with tools=[BoardDataFetcherTool(), CardDataFetcherTool()]
  • analysis_agent
• Tasks
  • data_collection
  • data_analysis
  • report_generation
• Crew

The analysis_agent used for two different tasks (data_analysis, report_generation) in this project

---

Complex crew Setups

---

Agentic Sales Pipeline (L6)

The Agentic Crew to create a Sales Pipeline was assembled using 2 Crews

Crew 1: lead_scoring_crew with 3 Agents, and 3 Tasks, and 2 Tools.

• Agents
  • lead_data_agent
  • cultural_fit_agent
  • scoring_validation_agent
• Tasks
  • lead_data_task
  • cultural_fit_task
  • scoring_validation_task with context=[lead_data_task, cultural_fit_task] and output_pydantic=LeadScoringResult`

Crew 2: email_writing_crew with 2 Agents, and 2 Tasks

• Agents
  • email_content_specialist
  • engagement_strategist
• Tasks
  • email_drafting
  • engagement_optimization

SalesPipeline - Built using crewai’s Flow feature

• Research a potential lead
• Score them given the person and the company details
• If it’s a qualified lead, then draft a proper initial email

Using crewai Flow, we can create end-to-end pipelines by leveraging Flow features like start, listen, and_, or_, router as well as state.

CrewAI operates on a higher level of abstraction, centered around the concept of a “crew” of autonomous AI agents. The “flow” in CrewAI dictates the sequence in which these agents, each with a specific role, set of tools, and a defined task, are activated. Key aspects include

• Role-based task delegation
• Sequential and Hierarchical workflows
• Implicit state management

This approach simplifies the process of creating multi-agent systems by focusing on the “who” and “what” rather than the “how” of the workflow. It excels in scenarios where the division of labor is clear and the process can be broken down into a series of distinct, role-based steps.

---

Performance Optimization

Speed and Quality are two main variables. However, Consistency is paramount.

How can we test agents performance?

By comparing the Task description with expected outcome using crewai test.

How can we improve the Task, Agent and the Crew without having to spend long time?

Leverage crewai train feature.