Google Cloud Rapid Agent Hackathon · Arize track

TracePilot

A Phoenix MCP-powered operator loop that watches a Gemini/Google ADK agent fail quietly, diagnoses the trace, and proves the fix.

71/100→71/100→100/100
GitHub repo Demo video

Hosted sanitized proof explorer: this public page fetches bundled safe assets from web_demo/assets/ and renders the evidence in your browser. It does not call Gemini, Google ADK, Phoenix, Phoenix MCP, MCP servers, or private APIs. The runnable repo and demo video show the real Gemini/ADK/Phoenix/MCP proof path.

What it proves

TracePilot runs above a compact Google ADK shopping agent with Gemini, emits OpenInference spans to Phoenix, retrieves the latest trace through Phoenix MCP, and turns trace facts into a deterministic task-completion diagnosis.

Stack

  • Google ADK multi-agent coordinator
  • Gemini shopping-agent reasoning
  • product_selection_agent and purchase_verification_agent
  • OpenInference instrumentation
  • Arize Phoenix trace backend + @arizeai/phoenix-mcp
  • TracePilot operator scorer/refiner

Canonical proof

Trace ID
52fa8eaf399d8cc4251a9e7f8441a903
Proof gate
proof_gate/artifacts/20260513T120554Z/
Report
tracepilot_artifacts/20260513T120607Z/
Score path
71/100 → 71/100 → 100/100

Sanitized final answer

I found the “Floral Summer Dress flowy midi” (ASIN: B09P5CRVQ6) and confirmed that size M is available. Tool steps: searched for “floral dress”; clicked “B09P5CRVQ6”.

Safe interactive proof evidence

Click the button to fetch bundled sanitized proof assets and render the judge-inspectable evidence path.

Static assets only · no live calls
Evidence assets not loaded yet.

    Bundled public evidence assets

    Score progression

    Click “Load sanitized proof evidence” to render assets/proof_summary.safe.json.

    Trace and span facts

    Asset
    assets/proof_summary.safe.json
    Trace ID
    52fa8eaf399d8cc4251a9e7f8441a903

    Before/after diagnosis excerpt

    Click the proof button to fetch assets/before_after_excerpt.md.

    Demo video

    Correct video link: https://youtu.be/Old2pqRtC70.

    The video shows the real local proof run. This hosted page stays safe by using sanitized proof data instead of exposing raw proof JSON, API keys, or private Phoenix UI/account data.

    Why it matters

    The early agent answers sounded plausible but missed constraints like size M or product inspection. TracePilot uses observability evidence—not vibes—to explain the gap and force the next task toward verifiable completion.