Agentic AI Testing / Scenario

Detecting Planning Failures at Scale

42,000 runs a week. Nobody reads traces at that volume. Your job: turn loop, stall, and drift pathologies into automated detectors with alerts.

Difficulty
Hard
Format
Scenario
Points
200
Estimate
14 min

// MISSION BRIEF

Your Mission

Your company's research agent handles 42,000 runs weekly. Support says users complain it "takes forever then gives up", cost is up 22%, and the only monitoring is a service-uptime ping.

You cannot read 42,000 traces. Design the detectors that find planning failures automatically, and decide what happens when they fire.

// FIRST CONTACT

Battle teaser

First artifact

Weekly ops digest

Which automated detectors should run over production traces? (Select all that apply.)

  1. AForced-stop rate as a first-class failure metric with an alert threshold
  2. BHire a reviewer to read every trace manually
  3. CStep-count outlier detector: runs far above the baseline for their task category
  4. DGoal-relevance heuristic: flag tool calls unrelated to the request category (get_weather in a supplier search)
  5. ENear-duplicate call detector: N consecutive calls to the same tool with trivially rephrased arguments and overlapping results
Answers, scoring, hints, and the full battle stay sealed.

// SKILL TAGS

agenticplanningmonitoringproductiondetection