Loading summary...
7 min read (71% time saved)
This video provides a comprehensive guide to building, productizing, and scaling vertical AI agents, contrasting them with horizontal agents and detailing a step-by-step playbook covering business models, pricing, finding opportunities, and practical implementation using an open-source framework and evaluation methods for continuous improvement.
The video addresses the need for a practical, step-by-step guide to building vertical AI agents, following up on a popular previous video. It promises to share a playbook for productizing and scaling AI agents, using a customer support agent for a SaaS platform as a real-world example.
Building vertical AI agents allows experts to multiply their impact by 100x across an industry, as their expertise is encapsulated in an agent that doesn't require direct involvement. This approach is also simpler than traditional SaaS, as it focuses on the agent's core functionality without complex UIs or backends, often integrating with existing communication platforms like Slack.
Vertical AI agents are built for specific roles and are reusable across multiple businesses, with the developer owning the intellectual property (IP) and charging for usage. In contrast, horizontal AI agents are custom-built for each business, with the client owning the IP, and developers typically charge hourly or per service. This distinction highlights the product-based nature of vertical agents versus the service-based nature of horizontal agents.
This section elaborates on the fundamental differences between product-based (vertical agents) and service-based (horizontal agents) business models, particularly in terms of scalability and upfront investment. Horizontal agents require less upfront investment but are less scalable, while vertical agents demand initial time and effort but offer significant scalability once productized. The speaker advises productizing as early as possible, even after just one or two agents, due to a market shift where clients prefer ready-made solutions over custom builds.
Three main strategies are presented for identifying valuable productization opportunities:
The most effective pricing model for vertical AI agents currently is licensing, where clients pay a recurring monthly fee for the agent's IP. Clients are also responsible for their own API token usage, which prevents unexpected cost spikes for the developer. Additional offerings include ongoing Slack support, onboarding, and a 7-day free trial, which is crucial for converting B2B SaaS customers.
Vertical AI agents are essentially templates. The building process involves identifying and productizing the parts of an agent that require customization across different businesses (e.g., instructions, knowledge, actions) while keeping the core functionality and general process consistent.
A seven-step playbook for building and scaling vertical AI agents is outlined:
The demonstration begins by using a starter template and the productize command within Cursor AI. The speaker explains the importance of customization, noting that while general processes are similar (e.g., accounting), specific steps and software vary. The framework handles this by using a custom onboarding tool, a Pydantic model, where fields are created for variables that change across businesses. An example of a data analytics agent's instruction template is shown, illustrating how variables are inserted to customize responses while maintaining a consistent process. The process of copying the template repository and opening it in Cursor is detailed.
The speaker proceeds to productize an existing customer support agent that handles about 30% of incoming tickets for their platform. Key customizable elements identified include the customer support email address, the customer support schema (for submitting requests), and the knowledge base (previous resolved issues). Using Cursor's agent mode and the productize command, the system suggests customizable fields. The speaker manually adds a missing customer_support_schemas variable and an environment variable for API authentication (customer_support_bearer_token) to the .env.template file. After testing the agent locally, it's pushed to GitHub for deployment on the Agencii platform. The deployment process on Agencii.ai is shown, including publishing to the marketplace (or keeping it private), adding a category, and optionally company details. The onboarding form, dynamically generated from the defined variables, is then used to customize and deploy the agent. The platform automatically updates client agents when changes are pushed and only requires new fields if the onboarding form is modified. The deployed agent is then integrated into a website widget, demonstrating its ability to search provided files, ask for details, and execute tools to submit a support request.
Evaluations (evals) are crucial for supercharging agents and creating a strong competitive advantage. The framework enables evals by default, accessible via OpenAI's platform under dashboard logs traces. Users can analyze messages, tool calls, and responses. To run an eval, one selects the agent, describes the desired check (e.g., if an issue was resolved correctly), chooses a model (GPT-5 mini recommended), and grades the traces. Evals run asynchronously, providing a resolution correctness percentage. The example shows a 60% correctness, with failed traces indicating areas for improvement, such as an agent correctly following escalation but not completing resolution within the trace. This continuous analysis of successful and failing traces allows for significant performance improvements and insights into customer problems.
The video concludes by reiterating the value of the tutorial, emphasizing that the demonstrated process is what the speaker's team is actively using to productize solutions previously built for clients, offering them at scale with reduced costs. The speaker encourages viewers to watch a follow-up video on an advanced data analytics agent.
"If you train an AI agent on your expertise, you can instantly 100x your impact across the entire industry because it no longer requires your direct involvement to serve clients."
— [Speaker]
"When you're building horizontal AI agents, you are in a service-based business. While if you are building vertical AI agents, you are in a product based business."
— [Speaker]
"Today I recommend you productize as soon as possible. Meaning that even after one to two agents, you should already start productizing them."
— [Speaker]
"The easiest way is to productize your own expertise."
— [Speaker]
"The key thing to understand is that vertical EI agents are just templates."
— [Speaker]
"Evals can give your agents superpowers because they will be able to continuously improve with every new client."
— [Speaker]
Summarize another video
Press ⌘K to quickly paste a new URL
IndyDevDan
5 min read
David Ondrej
6 min read
IndyDevDan
5 min read
Sean Kochel
6 min read
Y Combinator
7 min read
IndyDevDan
6 min read
