Creativity · Agent Protocol

Arize Phoenix for Agent Tracing and Evals

Phoenix is Arize's open-source AI observability library, tightly integrated with the OpenTelemetry GenAI conventions and OpenInference. It lets teams trace agent executions across frameworks (LangChain, LlamaIndex, DSPy, CrewAI), cluster failures by embedding similarity, run LLM-as-judge evals on traces, and do it all on a laptop or a self-hosted cluster — no vendor lock-in.

Protocol facts

Sponsor
Arize AI
Status
stable
Spec
https://phoenix.arize.com/
Interop with
OpenTelemetry, OpenInference, LangChain, LlamaIndex, DSPy

Frequently asked questions

What is OpenInference?

OpenInference is an open specification — maintained alongside Phoenix — for how LLM and agent spans should be named and attributed in OpenTelemetry. It's the schema that makes cross-framework tracing interoperable.

Do I need Arize Cloud to use Phoenix?

No. Phoenix runs fully locally via `pip install arize-phoenix` or as a self-hosted Docker container. Arize's paid cloud offers production-scale storage and team features.

Can Phoenix run evals?

Yes — it ships with hallucination, relevance, toxicity, and QA-correctness LLM-as-judge templates, and you can define your own.

Sources

  1. Arize Phoenix docs — accessed 2026-04-20
  2. Phoenix GitHub — accessed 2026-04-20