Agentic Al in SW Development: Evolving Patterns & Protocols • Bhuvaneswari Subramani • GOTO 2025

Bhuvaneswari Subramani presents a mental model for understanding the evolution of AI systems, which she terms “The Agentic Shift.” The presentation maps the landscape of AI system design by walking through seven foundational patterns, each overcoming the limitations of its predecessor. This journey illustrates the progression from simple AI helpers to complex, autonomous, and collaborative agents.

Framing the discussion with Andrej Karpathy’s vision of LLMs as a new kind of operating system, the talk details the following evolutionary patterns:

1. Conversational Clients: The most basic pattern, a simple, stateless loop where a user asks a question and an LLM provides an answer from its pre-trained data. This pattern lacks memory and context about the conversation.

2. Contextual Conversational Clients: This pattern enhances the basic model by providing a small amount of specific data (e.g., a document) as context along with the prompt. The LLM’s response is grounded in this provided context, but it’s limited by the size of the data that can be sent in a single prompt.

3. Retrieval-Augmented Generation (RAG): To overcome context size limitations, RAG indexes large volumes of enterprise data (PDFs, Confluence pages, codebases) into a vector store. When a user asks a question, the system retrieves relevant information from the vector store and provides it to the LLM, which then generates a response based on this private, proprietary data.

4. Self-Correcting RAG: This pattern adds a layer of validation and quality control to RAG. It implements a workflow (e.g., using LangGraph) where an LLM acts as a “judge” at multiple stages. The process involves checking the input prompt for toxicity, retrieving content, grading the relevance of retrieved documents, generating a final answer, and checking the answer for hallucinations before returning it to the user.

5. Function Calling Agents: This pattern enables the LLM to interact with the external world by calling predefined functions or tools. The LLM can choose the appropriate function to execute based on the user’s request. However, this approach can become cumbersome and difficult to scale as the number of available functions grows, often leading to complex conditional logic.

6. Agents with Model Context Protocol (MCP): MCP provides a standardized and scalable solution for tool use. An MCP server acts as a registry for all available tools (e.g., Jira, Git, Slack). The LLM is given the list of available tools and can dynamically select and invoke the correct one via the MCP server, eliminating hardcoded logic and making the system more modular. A demo showed an agent in the Kiro IDE using MCP to interact with a Jira instance.

7. Agent-to-Agent Interaction: The most advanced pattern, where multiple specialized agents, each potentially built on MCP, can communicate and collaborate. This allows for the creation of complex systems where different agents handle different tasks (e.g., one for planning, one for coding, one for testing) and work together to achieve a common goal.

• Tags:
• Agentic AI
• LLM
• RAG
• Feature
• Model Context Protocol
• MCP
• Retrieval Augmented Generation
• LangGraph
• Function Calling
• AI System Design
• AI Patterns
• Self-Correcting RAG