AI features on the web fail in ways console.error never sees: bad tool calls, runaway agent loops, slow model steps, and streams that die mid-flight. This talk walks through building agent experiences on the web and using Sentry agent tracing to observe full agent workflows in production. Then we bring in Seer, Sentry's own AI debugging agent, to turn that telemetry into root causes and fix guidance.

Agentic React apps where UIs autonomously act, chain tool calls, and evolve state—bring complexity that traditional observability tools can't unravel. How do you actually understand what your agents are doing in production?

In this talk, we'll explore agent observability: capturing not just logs, metrics, and traces, but the reasoning, tool usage, and decision paths behind agent behavior. You'll learn how to bring visibility into agent workflows, how OpenTelemetry semantic conventions help standardize telemetry, and how modern tools let your agents "talk back," so you can debug, understand, and trust what they're doing.

By the end, you will take away a practical workflow to instrument, trace, and interrogate agents—so your apps aren't operating in the dark.

AI agents are everywhere, but how do you go from demo to dependable? In this talk, we’ll build agents using LangGraph, and take it beyond toy examples into enterprise-grade readiness.

We’ll start with context engineering concepts like: control flow, prompt engineering, prompt management and tool calling. From there, we’ll build an agent step-by-step, enabling contextual reasoning via Retrieval-Augmented Generation (RAG), and integrating tool use for dynamic task execution. We'll also cover agent observability with open source observability tools such as LangFuse to bring full observability into the decision our agents make.

Whether you're exploring AI agents or planning your first production deployment, you'll leave this session with the patterns, tools, and clarity to do it right.

We want our software to be easy for people to use. That means building a great UI. Now that people work with AI agents, we need our software to be easy for agents to use. That means building a great MCP. Model Context Protocol servers give AI agents fingers to feel around and act in the world, one tool at a time.

In this session, I’ll share design considerations from the Honeycomb.io MCP. I’ll contrast MCP design with human UX and software APIs, then discuss ways to tell whether the MCP is effective in production.