Phoenix

Phoenix is an open-source Agent development and evaluation platform by Arize: visualize every step of the calls, prove the effectiveness of changes through experiments, start locally with one command, support self-hosting, and provide a solid basis for improving Agent quality.

AI Developer Tools#Open Source#Observability#Tracking

Updated: 2026-10-26

What is Phoenix?

Phoenix is an open-source platform launched by Arize, with the official slogan "Track Exponential Growth" and positioned as an open platform for Agent development and evaluation. Its core idea is simple: without tracing, you are flying blind. In addition to visualizing each step of the Agent's calls, it emphasizes using experiments to verify whether modifications truly improve quality. Officially, it is said to be used jointly by AI engineers and Fortune 500 companies. For Agent teams that pursue "letting data speak for themselves", it weaves tracing, evaluation and experiments into a workflow instead of a collection of scattered tools.

Who is it for?

Developers who iterate on Agent quality and need to quickly compare the effects of different versions; teams that want to run locally before deploying to the cloud; engineering organizations that need full visibility into production Agents; and teams that need to prove system quality to customers or compliance parties—persistent traces serve as ready-made evidence.

Key features

Distributed Tracing

Visualizes every step of calls made by Agents and LLM applications, enabling direct pinpointing of the exact step where an error occurs, eliminating guesswork.

PXI Interaction

You can directly converse, add annotations, and run experiments on a specific trace, turning the debugging process into a conversation for greater efficiency.

Evaluation and Experimentation

Systematically measure Agent quality, judge whether modifications are effective based on evidence rather than intuition, and form an improvement closed loop.

Local Startup

Run the command `uvx arize-phoenix serve` to start it locally with just one line, get started in seconds, and no complex configuration required.

Self-hostable

The open-source version can be deployed in your own environment. The official states that it can be deployed anywhere in seconds, giving you full control over your data and tech stack.

Agent-Native

Designed for Agent scenarios, not a simple port of traditional APM, capable of keeping up with the complexity of multi-step calls.

Experiment Comparison

Run multiple versions or models for the same problem, intuitively compare differences in output quality, and support decision-making on modifications.

Multi-framework Compatibility

Supports integration with mainstream LLM and Agent frameworks, with diverse access paths.

How to use Phoenix

The experience of starting with a single local command is ideal for the evaluation phase.
It is recommended to make "reviewing the trace first and then revising the prompts" a fixed workflow.
otherwise, it's easy to fall into a cycle of adjustments based on intuition.
Archive the evaluation results of each round of experiments, and after several rounds, you will be able to identify which modifications are actually effective.
For team collaboration, it is advisable to agree on a unified evaluation dataset and scoring criteria.
otherwise, the results developed by different members cannot be compared horizontally.
The official page also includes usage reviews from multiple practitioners, which are worth reading.

Things to know

The local version mainly serves development and evaluation; for production-level monitoring, team collaboration and hosting capabilities, refer to the boundaries of the open-source version and the service version.

Self-hosting requires you to maintain storage and handle upgrades yourself; when the volume of trace data is large, pay attention to storage and query performance planning.

The validity of evaluations depends on the evaluation criteria and datasets you design; if the metrics are poorly defined, the conclusions will be biased. Don't just focus on the good looks of the interface.

The platform is updated rapidly; keep track of documentation and versions to avoid using outdated APIs.

The local standalone version is suitable for evaluation, while for ultra-large-scale production environments, performance and capacity planning need to be considered.