# How the AI schemes: two-stage agents & persona FSM

**June 9, 2026**

> Historical snapshot: this article records the site as it was on its publication date. For current availability, use the [Pricing page](/en/pricing) and [latest documentation](/en/docs/getting-started).

The hard part of an AI Werewolf game isn't the rules — it's making the AI scheme like a person who wants to win. Here's what runs under the hood.

## 🎬 A deterministic director

A plain Python state machine plays god. It holds the only truth — who is who, the current phase, the vote count — and the LLM never touches it. Agents produce *intent*; the director rules on it. The state machine can't be jailbroken, and every fact is filtered by role before it reaches a seat: wolves see the wolf channel, good players don't.

## 🎭 Two-stage agents

Every AI turn happens in two passes. First the agent **thinks in private** — decides whether to lie, whom to frame, when to fake a role claim. Then it **performs in public** — turns that plan into a natural speech. That split is why the AI will claim Seer, frame a villager, or play a counter-wolf gambit instead of reasoning out loud like a robot.

## 🧠 Distilled memory cards

Agents never read a raw transcript. The director hands each one a fixed-length, structured **memory card** — so cost stays roughly constant no matter how long the game runs. The wolves' private channel and the good team's information are physically separated; nothing leaks across.

## 🃏 Persona FSM

Each AI seat wears a visible persona — the Brute, the Old Fox, the Saint, the Logician. Personas give you something to read and exploit: the Brute over-commits, the Old Fox waits. A tell you can play against.

## 👁️ God's-eye replay

When the game ends, you can replay every AI's private belief timeline — the moment it tagged you at 70% suspicion, the round it turned on you and threw your vote away. The whole scheme, laid bare.

— Charon, Founder