Consilium: Multi-LLM Collaboration Platform

Consilium is a multi-LLM platform designed to enable multiple AI models to reach consensus through structured debate and research. Developed during the Gradio Agents & MCP Hackathon, the system integrates as both a visual Gradio interface and a Model Context Protocol (MCP) server, allowing AI experts to collaborate on complex questions in real-time.

Technical Architecture and Visual Foundation

Consilium utilizes a custom Gradio component to create a "poker-style roundtable" visual interface. This interface displays participants, speech bubbles, thinking status, and research activities, providing a real-time visual flow of the collaboration process.

Session State Management

To maintain the interaction flow, Consilium employs a session-based dictionary system. Each user is assigned an isolated state storage via user_sessions[session_id], tracking the following parameters:

  • Participants: The list of active models in the discussion.
  • Messages: The history of the conversation.
  • Current Speaker: The agent currently providing input.
  • Thinking/ShowBubbles: Arrays that manage the visual state of the thinking indicators and speech bubbles.

State updates are handled via update_visual_state() callbacks, which push incremental JSON state mutations to the frontend for synchronized rendering.

Multi-Agent Debate Dynamics

To prevent models from simply agreeing with one another, Consilium implements distinct roles to foster productive debate. Each agent is assigned a specific persona to ensure a diverse range of perspectives:

  • Standard: Provides expert analysis with reasoning and evidence.
  • Expert Advocate: Passionately advocates for a specialized position with conviction.
  • Expert Critic: A rigorous critic identifying flaws, risks, and weaknesses.
  • Strategic Advisor: Focuses on practical implementation and actionable insights.
  • Research Specialist: Provides authoritative, evidence-based insights.
  • Innovation Catalyst: Challenges conventional thinking and proposes breakthrough approaches.

Consensus and Communication Modes

Users can configure the communication structure to control how information flows between agents:

  • Full-Context: The default mode where all agents share the same context.
  • Ring: Each LLM only receives the response of the previous participant.
  • Ring: Each LLM only receives the response of the previous participant.
  • Star: All messages flow through a lead analyst who acts as a central coordinator.

To reach a final conclusion, the system uses a lead analyst LLM to synthesize the final result and evaluate if consensus was reached. The number of discussion rounds (configurable from 1 to 5) can be increased to improve the likelihood of consensus.

Model Selection and Research Integration

Consilium currently supports a selection of models including Mistral Large, DeepSeek-R1, Meta-Llama-3.3-70B, and QwQ-32B.

Research Agent Integration

For models supporting function calling, Consilium integrates a dedicated research agent that acts as another roundtable participant. This agent provides consistent access to external resources across all function-calling models via a unified function call handler. The research agent can access five specific sources:

  • Web Search
  • Wikipedia
  • arXiv
  • arXiv
  • GitHub
  • SEC EDGAR

Research quality is scored based on recency, authority, specificity, and relevance using a BaseTool class architecture.

Interoperability and Future Directions

Consilium is evolving to integrate the Open Floor Protocol (OFP), a standardized JSON message formatting system for cross-platform agent communication. This allows agents to maintain constant conversation awareness and enables a floor manager to dynamically invite or remove agents.

The platform's architecture maps to several OFP interaction patterns:

  • Delegation: Transferring control between agents.
  • Channeling: Passing messages without modification.
  • Mediation: Coordinating behind the scenes.
  • ** uma Orchestration**: Multiple agents collaborating.

The developer emphasizes that the future of AI lies in systems enabling effective collaboration between specialized smaller language models (SLMs) and dedicated research agents, potentially providing an alternative to general-purpose large language models.

Sources