Arize AI

More than 40 integrations with models, frameworks and AI tools instrument the customer's application to send traces into Arize, code evaluators run in E2B, Daytona, Vercel or Modal sandboxes, and Data Fabric syncs trace data to warehouses. These move telemetry in and data out. No connectors that let an agent take authenticated actions in outside systems are documented.

SourceArize, arize.com homepage and pricing and the Phoenix README on github.com/Arize-airead 2026-09-21

Agents that Arize observes, evaluates and improves run elsewhere, and no workflows that sequence, branch or retry an agent's steps, or mix deterministic nodes with agent steps, are documented.

SourceArize, arize.com pricing and homepageread 2026-09-21

RAG pipelines built elsewhere are traced and evaluated in Arize, but no retrieval structure over the customer's own documents that grounds an agent's answers is documented.

SourceArize, arize.com homepage and pricingread 2026-09-21

Signal generates review-ready fixes for a person to review rather than applying them, so changes Arize's agents propose pass through human review. No approval step, escalation rule or pause before an agent acts at run time is documented, nor are autonomy modes that treat low risk and high risk actions differently.

SourceArize, arize.com homepage and pricingread 2026-09-21

Arize states it is certified to SOC 2 Type II, ISO 27001, PCI DSS, HIPAA and GDPR, with a Trust Center, and its plans document organization- and space-level RBAC, service accounts, Enterprise SSO and audit logs.

SourceArize, arize.com homepage and pricingread 2026-09-21

Agents are traced in Arize AX on OpenTelemetry with path and graph visualizations of each agent trajectory, token, latency and cost tracking, session and multi-modal tracing, custom dashboards, metrics, monitors and views on traces, audit logs on Enterprise, retention by plan (15 days Free, 30 days Pro, custom Enterprise), and Data Fabric sync of trace data to BigQuery, Databricks or Snowflake.

SourceArize, arize.com pricing and homepageread 2026-09-21

For tracing and evaluation, Arize stores agent trajectories and sessions, and no session, conversation or long-term memory that an agent reads back as context is documented.

SourceArize, arize.com homepage and pricingread 2026-09-21

Every Arize AX plan chooses a US, EU or CA data region, Enterprise adds self-hosted deployments, and open-source Phoenix runs locally, in containers with Docker and Helm, or in the customer's cloud.

SourceArize, arize.com pricing and the Phoenix README on github.com/Arize-airead 2026-09-21

Ready-made agents a buyer switches on include Alyx, an AI engineering agent that runs evals, debugs issues and improves the customer's agents, and Signal, whose agents find production issues and propose fixes, each working without the other, alongside managed agents on Enterprise and packaged skills that teach coding agents to trace, evaluate and debug with Phoenix. Those are role-specific agents and ready-made workflows.

SourceArize, arize.com homepage and pricing and the Phoenix README on github.com/Arize-airead 2026-09-21

Online evals run on traces and spans as they arrive, custom monitors fire on production conditions, and Signal's agents surface failure patterns from production telemetry, trace root causes and generate proposed fixes, so Arize's work starts from incoming telemetry with no person initiating each run.

SourceArize, arize.com homepage and pricingread 2026-09-21

The AI provider behind Arize's own agents and evaluators can be customized, with the customer's own key alongside Arize-managed models on Enterprise, and the Phoenix playground compares models and replays traced LLM calls against them. Customers control the models Arize's own AI features run on, though bringing their own key requires the Enterprise tier.

SourceArize, arize.com pricing and the Phoenix README on github.com/Arize-airead 2026-09-21

Arize publishes Python and TypeScript client, OpenTelemetry and evals packages, a GraphQL API, a CLI that fetches traces, datasets and experiments, and a remote MCP server at each Phoenix instance's /mcp endpoint that MCP clients use to query traces, datasets and experiments, all on the OpenInference and OpenTelemetry standards it maintains. That is a stable API and SDK for Arize's own platform, with MCP served.

SourceArize, the Phoenix README on github.com/Arize-ai and arize.com homepageread 2026-09-21

Testing covers online evals on traces and spans and offline evals on datasets and experiments, with trace evals of agent trajectories, session evals of multi-turn conversations, agent-as-a-judge and custom code evaluators, evaluator alignment, human annotations and labeling queues, agent experimentation in the playground, and Prompt Learning optimization; Phoenix adds open-source evals, datasets and experiments. That tests the customer's agent against datasets before production and scores quality over time.

SourceArize, arize.com pricing and the Phoenix README on github.com/Arize-airead 2026-09-21

Arize observes, evaluates and improves agents, and no browser, desktop or computer control by an agent is documented.

SourceArize, arize.com homepageread 2026-09-21