Overview

This session was a fast tour of how agentic AI app development is evolving in the wild. Julia Shottenstein (LangChain) and Michele Catasta (Replit) shared real deployments—Etsy’s Gift Mode “concierge” search, Home Depot’s Magic Apron that pairs products with how-to guidance, and a Replit customer whose app self-improves from user feedback. They compared notes on rapidly shipping (Replit Agent 1→3 in a year; LangChain 1.0 alpha) and why LangChain stays popular: it’s the quickest on-ramp to agents and lets teams swap models without rewrites, with LangSmith as a companion paid tool. Biggest build challenges: quality vs. reliability, plus cost/latency tradeoffs—solved with evals, observability, and tight product UX. Advice for devs: lean into “agent engineering” (context/prompt design, testing, online evals), get closer to users, and embrace AI tools. On security, they stressed opinionated scaffolds, authZ, and scanners alongside red-teaming. Risks: breakneck expectations; opportunities: better models, longer-lived autonomous agents, and a steadily improving craft.