Link to Folder with Everything

2025.03.19.Kono.Office.Hours.18 - Google Drive

Recording of Video

[https://drive.google.com/file/d/1fsqymdLkjkR6EolEiSWWDyJaASNT23PI/view?usp=sharing](https://drive.google.com/file/d/1fsqymdLkjkR6EolEiSWWDyJaASNT23PI/view?usp=sharing)

Summary ( by Gemini )

Summary

Rahul Singh, Gaja Kannan, and Merziyah Poonawala discussed creating reproducible AI agent documentation using a standardized "AI playbook" centered around a Product Requirements Document (PRD), focusing on the agent manifest (system prompt, tool registry, memory), and leveraging no-code tools like Nan.  The meeting included a demonstration of building a "news learner" agent in Nan, highlighting iterative development and tool refinement, showcasing the use of ChatGPT for prompt engineering and cross-platform adaptability. Future workshops will focus on building agents for customer service and news aggregation, with potential for a white paper on the standardized agent manifest.

Details

• Meeting Introduction and Introductions: Rahul Singh began the meeting by explaining the concept of "vibe development" and introducing the topic of creating easily reproducible AI agent documentation akin to a playbook (00:00:00). Several attendees introduced themselves, including Gaja Kannan, a long-time acquaintance of Rahul Singh with experience in AI-assisted coding, and Merziyah Poonawala, a product manager attending for the first time (00:01:21). Rahul Singh stated his goal was for attendees to understand AI agent design and implementation (00:02:21).
• AI Agent Design and Creation: Rahul Singh described AI agents as going beyond chatbots by incorporating memory, goals, and external tool usage (00:05:52). They emphasized the accessibility of AI creation using no-code tools and the potential for significantly increased effectiveness through AI agent development. They highlighted the importance of developing a standardized process for creating multiple agents (00:03:25) (00:07:03).
• AI Playbook and Product Requirements Documents (PRDs): Rahul Singh introduced an AI playbook centered on a well-defined PRD (00:08:07). They explained that a clear PRD, coupled with AI assistance, enables the creation and maintenance of sustainable software and intelligence systems, regardless of the chosen coding approach (00:09:26). The key to unique AI agents lies in their database schema and agent definition (00:10:32).
• Agent Manifest Structure: The core components of an agent manifest were identified as the system prompt, tool registry, and memory. The tool registry includes tool signatures (inputs and outputs) and descriptions (00:12:30) (00:15:10). Rahul Singh demonstrated iterative refinement of the agent manifest using ChatGPT, emphasizing the importance of detailed instructions and context for optimal results .
• ChatGPT Prompt Engineering for Agent Design: Rahul Singh employed a detailed ChatGPT prompt to generate an agent manifest for a "bill payment assistant" (00:16:24). They iteratively refined the prompt to increase the level of detail in the generated manifest, aiming for a level of detail suitable for direct implementation in no-code tools (00:19:11).
• Agent Manifest and Tool Utilization: The meeting showcased how a detailed agent manifest can facilitate the creation of agents within no-code platforms like Nan. Rahul Singh addressed how to integrate existing tools into the agent design process and emphasized the use of a standardized protocol like MCP for interoperability (00:31:47). The importance of clearly defining tools, ensuring their readiness, and managing the interaction layer was also highlighted.
• Building an AI Agent in Nan: Rahul Singh demonstrated the creation of a “news learner” agent within the Nan platform . They used a tool agent type, specifying a chat model (ChatGPT 4 0 mini) and adding tools for fetching news and information, including a manual tool for retrieving RSS feeds (00:36:23). They encountered challenges with using existing tools like Hacker News’s API but found a workaround with Wikipedia.
• Iterative Agent Development and Tool Refinement: Rahul Singh continued building the agent, adding tools for accessing different news sources (RSS, SERP API, Wikipedia) . They also demonstrated how to create a summarization tool within Nan, enhancing the agent's ability to process and present information concisely . The importance of iterative development and the ability to reuse and share these tools were noted.
• Agent Capabilities and Future Development: The demonstration showcased the agent's ability to gather information from multiple sources and answer complex questions by accessing and processing data from various tools (00:52:46). Rahul Singh concluded by discussing the potential for scheduling agent tasks and the application of this agent design process across different industries (00:54:16).
• Leveraging AI for Agent Building and Cross-Platform Adaptability: The meeting concluded with Rahul Singh highlighting the use of ChatGPT to generate step-by-step instructions for building the agent in various no-code platforms (00:56:34). This showcases how AI can assist not only in the design but also in adapting the implementation across different frameworks.
• Agent Creation Frameworks and No-Code Tools: Rahul Singh discussed various agent creation frameworks (LangChain, LangFlow, Flowise, Nan), highlighting Nan as an accessible, open-source option for non-programmers (01:00:01) (01:06:03). They emphasized the use of no-code tools like Nan and NoCodeDB for easier tool creation and maintenance (01:00:01), enabling non-programmers to build AI applications by assembling tools (01:02:32). The advantages of no-code databases (like NoCodeDB and Airtable) for data management and integration with agents were also discussed (01:01:28) (01:06:03).
• AI Playbook Components and Design: Rahul Singh presented an "AI playbook", outlining key components: agent manifest (system prompt), agent framework, and tool registry (describing tools and their input/output) (01:04:04) (01:07:16). They suggested starting simple with Nan and focusing on the design aspect (schema and flow diagrams) rather than the specific technology used. The choice of no-code versus code depends on individual preferences and project scale, with the core skill being the design and creation of agents, regardless of the chosen framework (01:13:21).