26 Key Takeaways from Building 150+ Agents in 9 months

0:00:00.000

1. Intro   (0:00:23.000)

  • Built over 150 AI agents in 9 months
  • Sharing 26 key takeaways
  • Aim to prevent others from repeating mistakes
  • 0:00:00.000

  • 1.1. Introduction to AI Agents   (0:00:23.000)
  • Overview of building AI agents
  • Purpose of sharing takeaways
  • Importance of learning from mistakes
  • 0:00:23.000

    2. Takeaway 1   (0:01:16.000)

  • AI agents differ from employees and automations
  • Agents require specific training
  • Think of agents in terms of Sops
  • 0:00:23.000

  • 2.1. Understanding AI Agents   (0:00:38.000)
  • Agents not employees or automations
  • Automations have hardcoded steps
  • Agents have less autonomy than employees
  • 0:01:01.000

  • 2.2. Training and Thinking About Agents   (0:00:38.000)
  • Agents need training on specific instructions
  • Cannot learn by trial and error
  • Consider agents in terms of Sops
  • 0:01:39.000

    3. Takeaway 2   (0:00:51.000)

  • Start with well-documented processes
  • Sops simplify agent training
  • Use existing onboarding materials
  • 0:01:39.000

  • 3.1. Importance of Sops   (0:00:51.000)
  • Begin with well-documented processes
  • Sops contain necessary training data
  • Utilize existing onboarding materials
  • 0:02:30.000

    4. Takeaway 3   (0:01:12.000)

  • Business owners won't build their own agents
  • Agent platforms will increase demand for developers
  • Determining which agents to build is crucial
  • 0:02:30.000

  • 4.1. Business Owners and Agent Development   (0:00:36.000)
  • Business owners won't build agents
  • Agent platforms will increase demand
  • Developers needed to determine agent needs
  • 0:03:06.000

  • 4.2. The Role of Agent Developers   (0:00:36.000)
  • Developers crucial for agent creation
  • Determining which agents to build is key
  • Platforms will not replace developers
  • 0:03:42.000

    5. Takeaway 4   (0:01:12.000)

  • Business owners often unsure of needed agents
  • Consulting is vital for identifying valuable agents
  • Start with customer journey mapping
  • 0:03:42.000

  • 5.1. Identifying Valuable Agents   (0:00:36.000)
  • Owners unsure of needed agents
  • Consulting identifies valuable agents
  • Start with customer journey mapping
  • 0:04:18.000

  • 5.2. Mapping Customer Journeys   (0:00:36.000)
  • Map customer journeys on Figma
  • Identify automation opportunities
  • Focus on valuable parts of the process
  • 0:04:54.000

    6. Takeaway 5   (0:01:22.000)

  • Start with as few agents as possible
  • Begin with one small agent
  • Add more agents after fine-tuning
  • 0:04:54.000

  • 6.1. Starting with Few Agents   (0:00:41.000)
  • Avoid building too many agents
  • Start with as few as possible
  • Begin with one small agent
  • 0:05:35.000

  • 6.2. Fine-Tuning and Expansion   (0:00:41.000)
  • Fine-tune the initial agent
  • Deploy and test with client
  • Add more agents as needed
  • 0:06:16.000

    7. Takeaway 6   (0:01:12.000)

  • Data-driven decisions are important
  • Combining data with actions yields results
  • Scrape both internal and external sources
  • 0:06:16.000

  • 7.1. Data-Driven Decisions   (0:00:36.000)
  • Importance of data-driven decisions
  • Garbage in, garbage out principle
  • Data quality affects agent outputs
  • 0:06:52.000

  • 7.2. Combining Data and Actions   (0:00:36.000)
  • Combine data with relevant actions
  • Achieve higher results with both
  • Scrape internal and external sources
  • 0:07:28.000

    8. Takeaway 7   (0:01:18.000)

  • Prompt engineering is an art
  • Provide examples and consider order
  • Iterate and test prompts constantly
  • 0:07:28.000

  • 8.1. The Art of Prompt Engineering   (0:00:39.000)
  • Prompt engineering is a real job
  • Importance grows with model evolution
  • Write prompts like essays or blog posts
  • 0:08:07.000

  • 8.2. Tips for Effective Prompts   (0:00:39.000)
  • Provide examples in prompts
  • Order of sentences matters
  • Iterate and test prompts constantly
  • 0:08:46.000

    9. Takeaway 8   (0:01:18.000)

  • Integrations are as important as functionality
  • Agents must work in existing systems
  • Convenience for users is crucial
  • 0:08:46.000

  • 9.1. Importance of Integrations   (0:00:39.000)
  • Integrations as important as functionality
  • Focus on agent capabilities often
  • Integrations affect user convenience
  • 0:09:25.000

  • 9.2. Integrating Agents into Systems   (0:00:39.000)
  • Integrate agents into existing systems
  • Example: customer support agent in Zendesk
  • Convenience drives value for users
  • 0:10:04.000

    10. Takeaway 9   (0:01:07.000)

  • Agent reliability has been solved
  • Use Pydantic for data validation
  • Developers are responsible for reliability
  • 0:10:04.000

  • 10.1. Solving Agent Reliability   (0:00:33.500)
  • Agent reliability issue resolved
  • Startups focus on reliability
  • Developer responsibility for reliability
  • 0:10:37.500

  • 10.2. Using Pydantic for Validation   (0:00:33.500)
  • Jason Liu's Pydantic solution
  • Validate agent inputs and outputs
  • Prevents major consequences
  • 0:11:11.000

    11. Takeaway 10   (0:00:54.000)

  • Tools are the most important component
  • Agents generate value through actions
  • Focus on building and structuring tools
  • 0:11:11.000

  • 11.1. Importance of Tools   (0:00:54.000)
  • Tools crucial for agent building
  • 70% of work goes into building actions
  • Tools provide value through actions
  • 0:12:05.000

    12. Takeaway 11   (0:00:54.000)

  • Limit tools to four to six per agent
  • Complexity affects tool limits
  • Splitting agents prevents hallucinations
  • 0:12:05.000

  • 12.1. Limiting Tools per Agent   (0:00:54.000)
  • Four to six tools per agent
  • Depends on tool complexity
  • Prevents agent hallucinations
  • 0:12:59.000

    13. Takeaway 12   (0:00:43.000)

  • Model costs are not a major concern
  • Focus on ROI rather than costs
  • Significant ROI from AI agents
  • 0:12:59.000

  • 13.1. Model Costs and ROI   (0:00:43.000)
  • Model costs not a major concern
  • Focus on ROI from AI agents
  • Example: significant cost reduction
  • 0:13:42.000

    14. Takeaway 13   (0:00:46.000)

  • Clients don't care about the model used
  • Use Azure OpenAI for data privacy
  • Developer experience with OpenAI API
  • 0:13:42.000

  • 14.1. Client Focus on Value   (0:00:46.000)
  • Clients focus on value, not models
  • Use Azure OpenAI for privacy
  • Developer experience with OpenAI API
  • 0:14:28.000

    15. Takeaway 14   (0:00:52.000)

  • Don't automate until value is established
  • Manual process validation is crucial
  • Automate after confirming value
  • 0:14:28.000

  • 15.1. Establishing Value Before Automation   (0:00:52.000)
  • Automate after establishing value
  • Manual process validation needed
  • Development costs are a concern
  • 0:15:20.000

    16. Takeaway 15   (0:01:03.000)

  • Focus on ROI, not just use cases
  • Use ROI formula to calculate value
  • Example calculation of ROI
  • 0:15:20.000

  • 16.1. Calculating ROI   (0:01:03.000)
  • Focus on ROI over use cases
  • Use ROI formula for calculations
  • Example: ROI calculation for a process
  • 0:16:23.000

    17. Takeaway 16   (0:00:50.000)

  • Agent development is iterative
  • Test different architectures
  • Compare and refine agent solutions
  • 0:16:23.000

  • 17.1. Iterative Agent Development   (0:00:50.000)
  • Agent development is iterative
  • Test various architectures
  • Compare and refine solutions
  • 0:17:13.000

    18. Takeaway 17   (0:01:03.000)

  • Use divide and conquer approach
  • Deliver solutions incrementally
  • Automate by departments first
  • 0:17:13.000

  • 18.1. Divide and Conquer Strategy   (0:01:03.000)
  • Use divide and conquer approach
  • Deliver solutions incrementally
  • Automate by departments first
  • 0:18:16.000

    19. Takeaway 18   (0:01:00.000)

  • Evals are important for big companies
  • Evals improve solutions over time
  • SMBs may not need evals initially
  • 0:18:16.000

  • 19.1. Importance of Evals   (0:01:00.000)
  • Evals crucial for big companies
  • Improve solutions over time
  • SMBs may not need evals initially
  • 0:19:16.000

    20. Takeaway 19   (0:01:11.000)

  • Two types of agents: agents and workflows
  • Workflows have predetermined steps
  • Combine workflows with agentic steps
  • 0:19:16.000

  • 20.1. Types of Agents   (0:01:11.000)
  • Agents and agentic workflows
  • Workflows have set steps
  • Combine workflows with agentic steps
  • 0:20:27.000

    21. Takeaway 20   (0:00:53.000)

  • Agents need to be adaptable on feedback
  • Add tools for analyzing results
  • Ensure agents can read and validate data
  • 0:20:27.000

  • 21.1. Adaptability and Feedback   (0:00:53.000)
  • Agents need adaptability on feedback
  • Add tools for result analysis
  • Ensure data reading and validation
  • 0:21:20.000

    22. Takeaway 21   (0:00:59.000)

  • Don't build around limitations
  • Models will continue to improve
  • Avoid building obvious general use cases
  • 0:21:20.000

  • 22.1. Building Beyond Limitations   (0:00:59.000)
  • Avoid building around limitations
  • Models will continue to improve
  • Don't focus on obvious general use cases
  • 0:22:19.000

    23. Takeaway 22   (0:00:57.000)

  • Deploying agents is harder than building
  • Integration into client processes is key
  • Building a platform for deployment
  • 0:22:19.000

  • 23.1. Challenges of Deploying Agents   (0:00:57.000)
  • Deployment harder than building
  • Integration into client processes
  • Building a platform for deployment
  • 0:23:16.000

    24. Takeaway 23   (0:00:55.000)

  • Waterfall projects don't work for agents
  • Use subscription-based models
  • Transition to agile service agreements
  • 0:23:16.000

  • 24.1. Agile Approach to Agent Projects   (0:00:55.000)
  • Waterfall projects unsuitable
  • Use subscription-based models
  • Transition to agile service agreements
  • 0:24:11.000

    25. Takeaway 24   (0:00:48.000)

  • Include humans in the loop for critical agents
  • Review and approve agent actions
  • Remove human loop after fine-tuning
  • 0:24:11.000

  • 25.1. Human in the Loop for Critical Agents   (0:00:48.000)
  • Include humans for critical agents
  • Review and approve agent actions
  • Remove human loop after fine-tuning
  • 0:24:59.000

    26. Takeaway 25   (0:00:53.000)

  • 2025 is the year of vertical AI agents
  • Vertical agents serve specific industries
  • Start with horizontal agents before vertical
  • 0:24:59.000

  • 26.1. Vertical AI Agents in 2025   (0:00:53.000)
  • 2025: year of vertical AI agents
  • Serve specific industries
  • Start with horizontal agents first
  • 0:25:52.000

    27. Takeaway 26   (0:00:56.000)

  • Agents help businesses scale
  • Automation leads to higher revenues
  • Employees focus on higher-level tasks
  • 0:25:52.000

  • 27.1. Agents and Business Scaling   (0:00:56.000)
  • Agents help businesses scale
  • Lead to higher revenues
  • Employees focus on higher-level tasks