Arize AX (Arize AI, 2020) — managed AI engineering platforma: Observe + Evaluate + Develop + Alyx + Signal + adb. Framework-agnostic, oparta na OpenInference + OpenTelemetry, 40+ integracji (OpenAI, Anthropic, LangGraph, LangChain, LlamaIndex, CrewAI). SaaS US/EU/CA lub self-hosted. Skala 1T span/mies., klienci Reddit, Uber, Spotify.

Arize AX to komercyjna, managed platforma AI engineering firmy Arize AI, pozycjonowana jako continual learning platform for agents — kompleksowy stack pokrywajacy caly cykl zycia aplikacji AI: obserwuj (Observe), oceniaj (Evaluate), rozwijaj (Develop), autonomicznie ulepszaj (Alyx / Signal). Zbudowana na OpenInference (wlasnym otwartym standardzie semantic conventions Arize) oraz OpenTelemetry — instrumentujesz raz, bez proprietary trace format. Framework agnostic: dziala z OpenAI, Anthropic, Google, Amazon Bedrock, LangGraph, LangChain, LlamaIndex, CrewAI, OpenAI Agents SDK, DSPy i 40+ innych stackow. Wersja open-source (Phoenix) dostepna na Apache 2.0.
Cztery filary produktu Arize AX. (1) Observe — OpenTelemetry-compliant tracing kazdego wywolania LLM/tool/agent, token/latency/cost tracking, session support, multi-modal tracing (image, voice, PDF), agent trajectory visualization (path i graph), custom dashboards/metrics/monitors/views, sync datalake z adb DataFabric. (2) Evaluate — online i offline evals na traces/spans/datasets/experiments (unlimited we wszystkich planach), custom code evaluators (Enterprise), Agent-as-a-Judge, Align evaluators, Trace Evals (agent trajectories), Session Evals (multi-turn conversations), multi-modal evaluation, human annotations i labeling queues.
(3) Develop — Playgrounds unlimited, versioning promptow i serving, trace replay w playground, multi-prompt comparison, agent experimentation, Prompt Learning (PL) optimization, datasety i experiments unlimited. (4) Improve (Alyx + Signal + Agent Swarms) — Alyx to in-app AI engineering agent (jak Cursor/Claude Code dla AI engineering) uruchamiajacy evals, debugujacy issues i improvujacy agenty klienta; Signal to always-on issue detection z auto-generation PRs; Agent Swarms orchestruja sandboxed managed debugging agents obsluguja swarm observability across managed i third-party agents. Enterprise dostaje Agent-as-a-Judge i dedicated engineers.
adb (Arize datastore) to purpose-built datastore dla GenAI traces — wg wlasnych benchmarkow najbardziej skalowana platforma do storage agent trajectories i kontekstu. Open formats, natywne polaczenie z BigQuery, Databricks, Snowflake via DataFabric. Skala operacyjna Arize: 1 bilion spans/miesiac, 1 miliard evals/miesiac, 5 milionow pobran Phoenix miesiecznie. adb i DataFabric dostepne od Enterprise tier.
Trzy plany + Startup program. AX Free (0 USD, 25k spans/mo, 1 GB storage, 15 dni retencji, unlimited users i evals, community support). AX Pro (50 USD/miesiac, 50k spans/mo, 10 GB storage, 30 dni retencji, email support). AX Enterprise (custom pricing, custom spans/storage/retention, SaaS lub self-hosted, RBAC, SSO, audit logs, HIPAA, SLA, Alyx, Signal, Agent Swarms, adb DataFabric, dedicated support). Startup pricing dostepny osobno.
Bezpieczenstwo klasy enterprise. Certyfikaty: SOC 2 Type II, ISO 27001, PCI DSS, HIPAA, GDPR (Trust Center: arize.com/trust-center). Data residency: US, EU, Kanada dla SaaS; self-hosted na AWS/GCP/Azure Kubernetes w wlasnym VPC klienta (Enterprise). Klienci: Reddit, DoorDash, Instacart, Uber, Spotify, Booking.com, TripAdvisor, PagerDuty, Snorkel, Handshake, Atlassian, PepsiCo, TheFork, Wayfair, Typeform, Priceline, Air Canada, Cohere, Conde Nast, BenchSci, BazaarVoice, Atropos, Defense Innovation Unit. Konkurencja: LangSmith (LangChain), Braintrust, DeepEval / Confident AI (open-source), Weights & Biases Weave, Fiddler AI, Langfuse. Wyroznik: OpenInference jako open standard + Phoenix (open-source, 5M+ pobran) buduja szerszy fundament niz konkurenci.