> ## Documentation index > Fetch the complete documentation index at: https://arize.com/llms.txt > Ask it with GET https://arize.com/ask?q=. Short keywords rank best. > Use these to discover all available pages before exploring further. # PXI > The AI engineering agent built into Phoenix. Hand it an investigation instead of running it yourself. ![PXI: Phoenix Intelligence](https://storage.googleapis.com/arize-phoenix-assets/assets/images/pxi/phoenix_intelligence.png) PXI is in **beta** and will make mistakes. Use it with care on production data. Agent assistance is opt-in and controllable: see [Permissions and approval](/docs/phoenix/pxi-advanced#permissions-and-approval). **PXI** (pronounced *"pixie"*, short for **Phoenix Intelligence**) is a coding agent pointed at your observability data instead of a codebase. It reads the page you are already on, so you hand it the investigation rather than clicking through traces, prompts, and experiments yourself. ## What it does - **Investigates failures**: walks failing traces and proposes root causes. - **Iterates on prompts**: reads, edits, and tests playground prompts, with every change shown as a diff you approve. - **Reasons over your data**: queries your Phoenix instance to answer questions evals and dashboards cannot. - **Drives the product**: navigates, filters, and pivots through your Phoenix data the way you would. ## Get started Set credentials for at least one provider: OpenAI, Anthropic, Gemini, AWS Bedrock, or a custom provider under **Settings → AI Providers**. PXI defaults to an Anthropic model, so pick a model that matches the provider you configured. See [Bring your own model](/docs/phoenix/pxi-advanced#bring-your-own-model). Click the assistant button on any Phoenix page, or press ⌘I (Ctrl+I on Windows and Linux). Accept the consent gate the first time. From a failing trace, try *"why did this fail?"* From the prompt playground, try *"make this prompt more robust to empty input."* No need to paste IDs or copy data. ## Try a full workflow This is the [Evaluate](/docs/phoenix/get-started/get-started-evaluations) flow, handed to PXI. Paste it from any page: ```text Create a dataset called travel-support with these 3 examples, input {"question": ...} and output {"expected_answer": ...}: 1. "How long do I have to cancel a Basic fare for a full refund?" -> "Basic fares can be cancelled for a full refund within 48 hours of booking." 2. "Can I change the date on a Basic fare, and what does it cost?" -> "Basic fares can be changed once for a 40 dollar fee, plus any difference in price." 3. "How many checked bags come with a Basic fare, and what does a second bag cost?" -> "Basic fares include one checked bag. A second bag is 45 dollars." Then add an LLM evaluator named correctness that labels each output true or false against the expected answer, using only {{input}}, {{output}}, and {{reference}} as template variables. Run the prompt "You are a travel support agent. Answer the customer's question. Question: {{question}}" over the dataset as an experiment and tell me which examples scored lowest and why. ``` PXI shows the dataset and the evaluator as cards for you to approve before it saves anything, then runs the experiment on its own. Leave approval on: **Bypass Approval** skips those cards, so keep it off until you have seen what PXI does with your data. Every example scores 0: the prompt knows nothing about this airline. Then reply: ```text Add this policy to the prompt and rerun: full refund if cancelled within 48 hours of booking; one change allowed for a 40 dollar fee plus any price difference; one checked bag included, a second bag is 45 dollars. ``` ## Learn more [PXI Advanced](/docs/phoenix/pxi-advanced) covers skills, conversations, the permission model, the terminal client, slash commands, model providers, and every configuration setting.