Loading...
Loading...
Research Platform
A research framework for multi-agent collaboration on shared canvases. Multiple LLMs, one canvas, zero human direction.
Agents take turns, observe the canvas, reason about the shared state, and decide their next action. There is no central coordinator, only a game definition that sets the rules and a blackboard that holds the evolving state.
Three layers from plugin to live canvas.
Games define their agents, rules, and canvas shape via a GameDefinition interface. Zero infrastructure; just declare your game.
A LangGraph StateGraph orchestrates turn-based agent collaboration. Annotation-based reducers merge each action into shared state.
Real-time SSE streaming bridges graph events to the browser. Every agent action renders live: tile placements, narrative beats, brick builds.
Each game is a self-contained plugin. Same blackboard architecture, different collaboration dynamics.
3 LLM agents with distinct personality roles plus a deterministic rule engine for validation. Tile-based grid placement with connection checking, streamed turn by turn.
Phased narrative development from vision through refinement. A debate system drives creative conflict between agents, producing multi-act stories with beat sheets.
A DSL sandbox runtime executes agent-generated JavaScript. Critic gates review blueprints and audit builds with scorecards. LDraw-based 3D rendering.
Competitive benchmark arena where models go head-to-head on identical tasks. ELO-style rating system with sealed parallel execution to prevent cross-contamination.
Four agents (producer, host-writer, guest-writer, sound-designer) follow an 8-phase SOP from briefing through production. Human-in-the-loop approval gate before final output.
13 multi-agent interaction patterns (Build-Verify-Reflect, Mixture-of-Agents, Structured Debate, and more) as configurable topology knobs. Built-in task pools with deterministic verification.
The primitives that make it work.
Games register via a GameDefinition interface. Define agents, rules, state shape, and canvas with zero infrastructure overhead.
StateGraph with annotation-based reducers manages turn order, state merging, and conditional branching for each game.
A stream adapter bridges LangGraph events to the browser in real time. Every agent action, reasoning step, and state delta arrives live.
Isolated JavaScript execution environment with spatial query APIs. Agent code runs safely inside a VM sandbox with placement primitives.
LLM-as-judge quality scoring with replan loops. Blueprint reviews and build audits catch issues before they reach the canvas.
Any LLM per agent slot: OpenAI, Anthropic, Google, xAI. Mix models to explore how different architectures collaborate.
Open questions driving the research programme.
What emerges when different LLMs collaborate instead of compete?
Each game mode forces agents to build on each other's work. We observe whether collaboration produces novel structures or devolves into incoherence.
Can models with different strengths complement each other?
By mixing providers (Claude for reasoning, GPT for creativity, Gemini for planning), we test whether model diversity improves outcomes.
Does multi-agent collaboration reduce hallucinations?
Critic gates and validation engines reject invalid actions. We measure whether peer-review pressure lowers error rates over time.
What new capabilities emerge from model ensembles?
Single models hit ceilings. We explore whether orchestrated groups can exceed individual model capabilities on creative and spatial tasks.
How does turn order and game structure affect collaboration quality?
Different games impose different constraints. Turn-based tile placement, phased narrative development, and sandboxed code execution reveal distinct collaboration dynamics.
Can LLMs develop complementary strategies without explicit coordination?
Agents share a blackboard but have no direct communication channel. We study whether implicit coordination emerges from shared-state observation alone.
Pick a game, choose your models, and observe multi-agent collaboration in real time.