AG2 - Agents for Production Engineering @arizeai

Updated September 2026, 2 weeks ago

Triaging Agent Errors with Phoenix and PXI

Why Most AI Agents Fail—and How Anthropic Builds Reliable Ones | Arize Observe 2026

$How to Build the Right Evals for AI Agents | Arize Phoenix AI agent evaluation can feel overwhelming. Agents use tools, skills, memory, files, and multi-turn conversations, creating far more failure points than a traditional LLM application. In this Arize:Observe 2026 session, Elizabeth Hutton, Senior Software Engineer for Evals at Arize, shares what the Phoenix team learned while building and evaluating Pixie, an AI engineering agent inside the open source Arize Phoenix platform. Rather than trying to design a complete evaluation suite upfront, the team built evals gradually as the agent gained new capabilities. Elizabeth explains how tracing, targeted capability evals, regression testing, production feedback, and trace analysis work together to create a practical agent improvement loop. The session covers: • Why tracing should come before formal evaluations • How traces become the source of truth for debugging agents • How to create small capability evals for tools, skills, and output formatting • Using synthetic data to cover edge cases, typos, and negative examples • Storing eval datasets and harnesses alongside application code • Running agent evals and experiments with coding agents • Turning targeted evals into regression checks in CI • Combining narrow evals with end-to-end Playwright tests • Why synthetic first-turn queries fail to represent real agent behavior • Building realistic multi-turn datasets from production traces • Using user feedback, LLM-as-a-judge, code evaluators, and trajectory evals • Reviewing traces to discover failure modes you did not anticipate • Moving production failures into development and regression datasets • Closing the loop between tracing, experiments, implementation, and evals The central lesson: evaluation is a development practice, not a one-time project. Start with tracing, add narrow evals as capabilities emerge, and expand the suite using failures discovered in production. :contentReference[oaicite:0]{index=0} Chapters: 00:00 Why AI agent evals feel overwhelming 01:09 Meet Pixie, Phoenix’s AI engineering agent 02:33 Agent architecture and the growing failure surface 03:34 Start with tracing 04:53 Add narrow capability evals 06:52 Turn capability evals into regression tests 08:53 Why synthetic evaluation data falls short 11:03 Discovery evals for production agents 12:04 User feedback and automated evaluations 12:58 Trace review and agent error analysis 14:33 Closing the agent improvement loop 14:59 Lessons learned from building Pixie 16:11 Open source evals in Arize Phoenix 🔗 Explore Arize Phoenix: https://phoenix.arize.com 🔗 View Phoenix on GitHub: https://github.com/Arize-ai/phoenix 🔗 Read the documentation: https://arize.com/docs/phoenix 🔔 Subscribe for more videos about AI agents, LLM evaluation, observability, and open source AI: https://www.youtube.com/@arizeai?sub_confirmation=1 #ArizePhoenix #LLMEvals #AIAgents$

Multi-Agent Frameworks: Building & Debugging with Groq and LlamaIndex

How to test AI agents with traces, evals, and CI/CD

I Told It to Pass the Tests... So It Deleted Them.

Introducing the New Arize Phoenix Open Source LLM Evals Library

Tracing Agents and Running Evals in TypeScript