The OpenTelemetry foundation is the differentiator. Because traces follow the shared OpenTelemetry for GenAI conventions rather than a proprietary format, Phoenix observes a stack through the same spans an application already emits, and its integrations directory covers every major framework (Phoenix integrations). Instrumentation is not locked to one agent framework’s idea of what a step looks like.
How does it fit an evaluation workflow?
Traces alone tell you what happened; evaluations tell you whether it was good enough. Phoenix turns trace data into datasets, runs them through LLM-based evaluators or human review, and tracks results as experiments across prompt or model changes. For a RAG application this means measuring retrieval quality per query, not just end-to-end vibes; for an agent it means scoring each tool decision inside a run rather than only the final answer. The same machinery powers regression detection in evaluation harnesses before changes reach production.
When is Phoenix the right choice?
When vendor neutrality or self-hosting drives the decision. Phoenix runs anywhere from a laptop to your own cluster under an open license, common in regulated environments and in stacks that avoid single-vendor lock-in. Compared with LangSmith or Langfuse, the capability set is similar; the deciding factors are licensing, hosting, and which standard your instrumentation speaks.