Agentic AI Testing / Scenario

The 95% Benchmark

A vendor claims 95% on their shiny agent benchmark. Read the fine print and decide how much of that number you should believe.

Difficulty
Hard
Format
Scenario
Points
200
Estimate
15 min

// MISSION BRIEF

Your Mission

A vendor pitches their coding agent: "95% success on our AgentBench-Pro suite." Your team is about to buy based on that number. Before you do, read how the benchmark was actually run.

Separate the real signal from the marketing.

// FIRST CONTACT

Battle teaser

First artifact

The vendor's benchmark methodology

Which methodology flaws inflate the vendor's 95%? (Select all that apply.)

  1. ASingle run with best-of-retries reported (hides non-determinism)
  2. BTiny toy tasks unlike production (weak construct validity)
  3. CSuccess judged by the agent's OWN generated tests (reward hacking / self-grading)
  4. DTasks published 14 months ago (likely training contamination)
Answers, scoring, hints, and the full battle stay sealed.

// SKILL TAGS

agenticbenchmarkscontaminationreliabilityevaluation