Rubber Duck Thursdays: Building an AI agent app

Key Concepts

• Agentic Loop: The iterative process AI agents use to gather context, take action via tools, and validate results.
• AI Agents: Autonomous software designed to perform specific tasks (e.g., sales, coding, 3D modeling/CAD) by interacting with environments and tools.
• LangChain: A framework for building AI-powered applications, supporting Python and JavaScript.
• GitHub Copilot SDK: A toolkit for integrating AI capabilities directly into applications.
• MCP (Model Context Protocol): A standard for connecting AI agents to external data sources and tools (e.g., PostgreSQL, Playwright).
• Entra ID: A secure identity management service used to handle authentication without exposing sensitive environment variables or API keys.
• Playwright: A tool used for browser automation, which can be integrated via MCP to allow agents to test their own UI interactions.

---

1. Main Topics and Key Points

The session focused on the development of AI agents, specifically the "agentic loop" and the practical implementation of agents within applications. Marlene, a maintainer for LangChain Azure, demonstrated how developers can move beyond pre-built tools like GitHub Copilot to build custom agents that automate business processes.

• The Agentic Loop: Agents operate by gathering context, executing tool calls, and performing self-validation.
• Security: A major emphasis was placed on avoiding hardcoded environment variables. The speaker advocates for Entra ID to manage credentials securely via tokens that refresh automatically.
• Frameworks: The speaker highlighted LangChain (for Python/JS) and the GitHub Copilot SDK as primary tools for building these agents.

2. Real-World Applications

• Sales Automation: A "Zava" sales concierge agent was demonstrated, which interacts with customers, checks pricing, and accesses case studies via a PostgreSQL database to answer queries.
• CAD Agents: The discussion touched on "Computer-Aided Design" agents, which allow users to generate 3D models and architectural designs through natural language prompts.
• AI Recruiters: Participants shared experiences building agents to automate recruitment workflows.
• Testing: Using the Playwright MCP server, agents can autonomously navigate a website, click buttons, and validate that the application functions correctly in real-time.

3. Methodologies and Frameworks

• Building an Agent:
  1. Define Intent: The agent uses an LLM to interpret the user's query.
  2. Tool Selection: The agent selects the appropriate tool (e.g., database query, calendar booking).
  3. Execution: The agent performs the task.
  4. Validation: Using tools like Playwright, the agent verifies the output.
• Integration: Developers can use create_agent functions in LangChain or create_session in the Copilot SDK to initialize the agentic environment.

4. Key Arguments and Perspectives

• AI as a Developer Tool: The speaker argues against the notion that AI will replace developers. Instead, she posits that AI creates a massive demand for developers to build specialized AI assistants for various industries.
• Self-Testing Agents: A significant takeaway is the importance of "machines testing machines." By allowing agents to validate their own work, developers reduce the manual burden of quality assurance.

5. Notable Quotes

• "One of the things is that we don't—sometimes the agents make mistakes and it's hard for you as a human to always go and validate that the agent has done things correctly, and so using something like this [Playwright]... is really fun." — Marlene
• "I don't think that [AI replacing developers] is really going to be true... because now that we know that AI is a thing, in lots of different industries, they're wanting us to build AI assistants for them." — Marlene

6. Technical Terms

• Middleware: A layer in the agentic framework that allows developers to run specific functions (like guardrails) before every LLM call.
• Bicep: An infrastructure-as-code language for Azure, mentioned in the context of available MCP servers for deployment.
• Async: Used in Python to ensure that agentic processes are efficient and non-blocking.

7. Synthesis and Conclusion

The session provided a practical look at the current state of AI agent development. The main takeaway is that building effective agents requires a combination of robust frameworks (LangChain/Copilot SDK), secure identity management (Entra ID), and autonomous validation (Playwright). By shifting from simple chatbots to agents that can interact with databases and UI environments, developers can automate complex, multi-step workflows. The speaker encourages developers to experiment with existing MCP servers to extend the capabilities of their agents.