Saurabh Dahal

Anyone can spin up an AI agent in five lines. Would you give that agent a refund button? In this demo-driven talk I build a customer support bot with LangGraph, the framework behind agents at Klarna, Uber, and J.P. Morgan, to show why control matters in production. Instead of letting the model improvise, I lay the bot out as a graph: it answers the easy questions on its own, pauses for human approval before anything irreversible like a refund, and resumes where it stopped after a crash. Then I ship it to AWS. Claude on Amazon Bedrock runs the model, Amazon Bedrock AgentCore hosts the bot, and the Agent Toolkit for AWS handles model access, IAM, and deployment from inside the coding agent. You'll leave knowing how to start with LangGraph, how to run it on AWS without a platform team, and when a graph beats a free-roaming bot.

As codebases grow, AI coding assistants struggle. Context windows overflow, agents lose track of dependencies, and simple text search fails to capture the semantic relationships that define real software. This talk explores three proven architectural patterns that enable AI agents to work effectively with production-scale codebases: semantic code intelligence through Language Server Protocol integration, specialized agent skills via context and tool bundling, and subagent delegation for efficient context management. Through live demonstrations on a popular open-source project like ShadCN, you'll see these techniques tackle the complexity of real-world software—multi-file refactorings,cross-module changes, and dependency tracking that would overwhelm traditional approaches.You'll leave with practical, product-agnostic strategies for building or enhancing AI agents that can handle large codebases with accuracy and efficiency. We'll examine why semantic understanding outperforms text search, how to design focused agent skills that improve task completion, and how parallel subagent architectures prevent context window exhaustion. Whether you're building AI developer tools, architecting multi-agent systems, or contributing to open source, these patterns will help you bridge the gap between toy demos and production-grade AI assistance.