Agent Trace Graphs & Decision-Level Visibility

Arize's agent trace graphs visualize the full internal state machine of agentic systems — tool calls, sub-agent delegation, retrieval steps, and decision branches — in a single interactive view. Unlike raw span logs that show you what happened, agent graphs show you why it happened, catching failure modes that look like success: unnecessary tool calls, wasted token loops, hallucinated arguments, and syntactically valid but semantically wrong outputs. The platform automatically detects recursive loops and repeated failures, flagging agent trajectories that consume budget without making progress [3].

Alyx — AI Engineering Agent

Alyx is an AI debugging assistant built into the Arize platform that functions like Cursor or Claude Code, but specifically for AI engineering. It runs evaluations, debugs trace failures, spots pattern issues in production data, optimizes prompts, and can even fix agent code. Give Alyx a problem trace, and it investigates the failure path, suggests root causes, and recommends fixes. This is a unique differentiator — no other observability platform ships an embedded AI agent for self-diagnosis [4].

MCP (Model Context Protocol) Tracing

Arize supports native tracing for Model Context Protocol, the emerging standard for connecting AI agents to external tools. MCP tracing captures every tool call, context fetch, and error response in the protocol exchange — giving developers visibility into how their agents interact with databases, APIs, filesystems, and other MCP servers. This is critical for production debugging because MCP errors often manifest as agent retry loops that silently burn tokens and degrade user experience [3].

Multi-Agent Graph Monitoring

For systems running multiple coordinated agents (e.g., a research agent delegating to a code agent that spawns a test agent), Arize renders the full multi-agent interaction graph. You can filter by agent, session, user, or time window to identify which agent in the chain is introducing errors, taking too long, or consuming disproportionate resources. Session-level evaluations measure end-to-end goal achievement across the entire multi-agent conversation [3].

OpenTelemetry-Native Instrumentation

Arize's architecture is built on OpenInference, the open-source leader in GenAI semantic conventions for OpenTelemetry. This means your tracing data is vendor-agnostic — the same instrumentation works with any OpenTelemetry-compatible backend. You can switch tools without re-instrumenting your code. The SDK-based approach (Python and JavaScript) is resilient: agents continue functioning even if the observability backend is down, unlike proxy-based alternatives that create a single point of failure [2]. Arize ships integrations with 40+ frameworks including OpenAI Agents SDK, Claude Agent SDK, LangGraph, CrewAI, AutoGen AgentChat, Pydantic AI, LlamaIndex, DSPy, Vercel AI SDK, and Google ADK.

Evaluation Framework

Arize's evaluation system supports span, trace, and session-level evaluations at scale — LLM-as-a-judge with configurable criteria, code-based evaluators for deterministic checks, and human annotation queues for expert review. The platform processes 1 billion evaluations monthly, running them both offline (on curated datasets) and online (on production traffic as it flows through). Evaluation results feed into dashboards, monitors, and the improvement loop — production failures can be promoted to test datasets with one click for regression testing [1].