Beyond Wrappers: The Council's Multi-Agent Architecture

Many demos claiming to be "Multi-Agent" simply have several LLMs with different prompts taking turns to speak. While this might sound lively, there is a lack of genuine interaction and confrontation between the roles, and the user is left merely as a spectator waiting for the result.

The Council's technical architecture was never designed to be just a simple chatbot wrapper. It was built to support a deeply confrontational exchange of ideas that can be interrupted and redirected. We built four core layers of technical barriers.

Layer 1: Dynamic Role Engine

The system doesn't use static role templates. When a user inputs a decision problem, the system converts it into a structured Decision Context. Based on this context, it selects from a preset role library or dynamically generates the most appropriate combination of roles.

• For career questions, the system might summon a "Career Planner," a "Financial Realist," and a "Devil's Advocate."
• For startup questions, it might bring in a "User Researcher," an "Investor Perspective," and a "Risk Controller."

Additionally, we introduced "Celebrity Debater" mental models and allow users to customize their own dedicated decision roles.

Layer 2: Stance Conflict Allocation System

If all roles strive for objective neutrality, the discussion becomes mediocre. The Council's middle layer is responsible for assigning clear stances and tasks to each Agent: support, oppose, supplement, cross-examine, or highlight risks.

This forced stance allocation ensures that every role has a unique entry point in the debate, avoiding the awkward situation where everyone "says pretty much the same thing."

Layer 3: Moderator Orchestration Layer

This is the core brain of The Council. If multiple Agents are left to speak freely, they can easily fall into infinite loops or go off-topic. The Moderator Agent is responsible for global orchestration:

1. Preventing Repetition: Monitors the content being spoken to prevent Agents from raising points that have already been discussed.
2. Probing Ambiguity: When an Agent's logic is loose, the Moderator will demand clarification.
3. Organizing Cross-Examination: Identifies core disagreements and guides Agents with opposing stances into direct confrontation.

Layer 4: Real-time Evidence Injection and Rollback Mechanism

This is The Council's killer feature. In real discussions, new information appears at any time.

When a user submits new evidence in the middle of a debate, the system immediately interrupts the current generation process. Through underlying state management, the system executes a Rollback, discarding the currently incomplete or outdated output, and forces the Agents to reorganize their thoughts incorporating the new evidence.

This reactive context management makes users truly feel they are participating in a real meeting, rather than watching a pre-recorded video.

Through this architecture, The Council ensures that each AI can think independently while working collaboratively, ultimately delivering a high-quality decision brief containing core disagreements and actionable advice for the user.