Coursera course

Multi AI Agent Systems with crewAI by DeepLearning.AI

Resumé and job seeking

4 Agents

1. Tech job researcher
2. Personal profiler for engineers
3. Resume strategist for engineers
4. Engineering interview preparer

Convert potential customers to customers

4 agents

1. research (data collection) on potential customer
2. comparison to other known customers
3. score this type of potential customer
4. create talking points

Mult-agent customer support agent

Follows the framework, define the following attributes to optimise performance:

1. Role playing, Focus, Cooperation
2. Tools, guardrails, Memory

Mental framework for agent creation

• Think as a manager
• Thank about what is the goal and what is the process
• What kind of people I would hire to get the job done?
• Use keywords to be specific about the essential qualifications

Specific to CrewAI

• Agents 'self improve" using memory
• Guardrails prevent agents from going into "rabbit holes" (unproductive, repetitive loops)
• Agents always attempt to get to an answer (avoid iterating indefinitely)
• Focus
  • narrowly defined task
  • specific agent roles and objectives
  • limited set of tools assigned to one agent

Customer outreach campaign

• Need tools that have the following attributes:
  1. Fault-tolerant (fail gracefully, not stop execution, send error message back to agent to retry)
  2. Versatile (handle different types of inputs)
  3. Caching (smart caching, cross-agent caching, prevent unnecessary caching, stay within rate limits, shorter execution times)

Key elements of well-defined tasks

1. Clear description of the task
2. Set a clear and concise expectation
3. Set context

Build a Crew to Tailor Job Applications (Lesson 7)

Goal: Maximise chances of getting an interview for a given job posting Process:

• Learn about the job requirements
• Cross that with your skill set and experiences
• Reshape resumé to highlight relevant areas
• Rewrite resumé with appropriate language
• Set talking points for your initial interview

5. Set a callback
6. Override Agent tools with specific task tools
7. Force human input before end of task
8. Execute asynchronously
9. Output as Pydantic / JSON / file
10. Run in parallel