Agentic Debate

Stress-test any idea, decision, code, or proposal using a 3-agent adversarial debate that exploits sycophancy via scoring incentives to produce high-fidelity analysis.

Based on the technique from "How To Be A World-Class Agentic Engineer" by @systematicls.

How It Works

LLMs are eager to please. Instead of fighting this, we exploit it:

• The Explorer is incentivized to find as many issues/angles as possible (scored per finding)
• The Adversary is incentivized to disprove findings (earns the finding's score) but penalized double for wrong disproves (-2x) — creating calibrated skepticism
• The Referee is told we have ground truth and scored on accuracy (+1 correct, -1 wrong) — pressured toward truth

The result is a superset of findings, filtered through adversarial challenge, judged by an accuracy-motivated referee.

Workflow

Step 1 — Interpret the Topic