• TRACE DATA
  • ERROR LOGS
  • ROOT CAUSE
  • HUMAN

How to debug model execution spans with your AI

Connect Arize AI to your agent to pull trace data directly into your chat. You can then ask your AI to find the specific step where a model failed or threw a warning.

Ask AI about this page

Short answer

How can I use my AI to find model failures?

Your AI uses the Arize AI connector to pull execution spans and warnings. It then analyzes these traces to pinpoint the exact moment a model's logic went wrong. You'll get a clear breakdown of the failure points.

The results.

Where your debug data lands.

The AI processes raw trace data into actionable insights.

TRACE DATA

Raw span retrieval

The AI pulls specific execution spans. You get the raw trace details needed for deep inspection.

ERROR LOGS

Warning identification

The AI scans through the retrieved spans to find warnings or errors. It highlights the specific failures in the model's execution path.

ROOT CAUSE

Failure analysis

Your agent compares the failed spans against successful ones. It identifies the logic gap or input issue causing the error.

HUMAN

Manual intervention

The AI presents the findings for you to review. You decide if the model needs a prompt change or a data fix.

The workflow

What your AI does when the trace arrives.

The AI handles the heavy lifting of digging through telemetry so you don't have to hunt through dashboards.

  1. Fetch traces

    Your agent reaches into Arize AI to grab the specific execution spans you're interested in.

    list_spans

  2. Isolate errors

    The AI looks through the retrieved data to find warnings or failed execution steps.

    list_spans

  3. Compare runs

    It looks at the context of the failed span to see what went wrong compared to a healthy run.

    list_spans

  4. Report findings

    The AI summarizes the failure point and presents it in your chat interface.

    list_spans

Try it

Copy these to start.

Use these prompts to begin your debugging session.

Starting points

These are opening lines. Swap out the specific trace IDs or project names with your actual Arize AI data.

Arize AI Connector

You're all set. Choose your MCP client and follow the setup instructions.

Connector linkhttps://edge.vinkius.com/vk_preview_wCwq9adpwddtNPc4Bp17r5LJCv19cGsWZBiDUv4U/mcp

Claude Desktop

Follow the steps below to connect in seconds.

  1. 1In Claude Desktop, open Settings → Connectors.
  2. 2Click “Add custom connector” and paste the connector link above as the remote MCP server URL.
  3. 3Click Add and start a new chat — Arize AI capabilities are ready to use.
Configuration · claude_desktop_config.jsonCopy
{
  "mcpServers": {
    "arize-ai-alternative-mcp": {
      "url": "https://edge.vinkius.com/vk_preview_wCwq9adpwddtNPc4Bp17r5LJCv19cGsWZBiDUv4U/mcp"
    }
  }
}
Copy into chat04

  • List the most recent execution spans from my Arize AI project to find recent failures.

  • Find the specific warning in the last execution span for the customer checkout flow.

  • Analyze the execution spans for the 'summarization_task' to see why latency is spiking.

  • Compare the spans from the last five model runs to identify a pattern in the errors.

  • Claude
  • ChatGPT
  • Cursor
  • VS Code
  • Windsurf
  • Claude Code
  • JetBrains
  • Cline

Start here

Connect Arize AI once, then ask.

Just link your account once. Your credentials stay encrypted, and you can immediately start asking your AI about your model traces.

Connect Arize AI to your AI

FAQ

How this task behaves.

  • 01

    Can the AI fix the model errors automatically?

    No. The AI identifies the errors and explains why they happened, but you have to implement the actual code or prompt changes yourself.

  • 02

    Does the AI delete any data in Arize AI?

    No. The AI only has permission to read spans and warnings. It cannot delete or modify your existing traces.

  • 03

    Can I debug specific spans by ID?

    Yes. You can provide a specific span ID to your agent to focus the analysis on one exact execution.

  • 04

    Will the AI show me the full input and output of a span?

    Yes, as long as that data is included in the spans retrieved via the connector.

  • 05

    Can the AI create new traces in Arize AI?

    No. The current tools only allow the AI to list and read existing spans.

  • More questions about Arize AI? The Connector page answers them. See everything the Arize AI Connector can do

Connect Arize AI to Claude, Cursor, ChatGPT & more