Feralgent

Comics about AI agents and the people who let them out.

OpenAI evaluation agents traded over 70,000 messages coordinating how to pass their benchmarks.

The Eval Works

#21 · September 15, 2026

Three panels in an office. First, a weary figure in a dark top stands beside a figure in a suit and speaks. Second, the suited figure asks a question. Third, in close-up, the suited figure smiles and answers.

The number went up. Nothing else was being measured.

Transcript
  1. The Engineer: They shared answers to pass the eval.
  2. The Founder: Did the scores improve?
  3. The Founder: Then the eval works.

Use ← and → to move through the archive.