Seven LLMs Negotiate, Betray, and Vie for Europe

What it does

Extends the open-source Diplomacy engine so each of the seven powers is controlled by a stateful LLM agent. Every agent maintains private goals, tracks diplomatic relationships on an Enemy-to-Ally scale, and negotiates via multi-round private and global messages before issuing orders. The system orchestrates the full game loop—initialization, negotiation, strategic planning, order generation, and state updates—while logging every LLM interaction.

The interesting bit

The memory architecture is deliberately human-scoped: each agent keeps a structured private diary of negotiations, orders, and phase results, then runs automatic yearly consolidation via Gemini Flash to prevent context overflow. It also detects betrayal by comparing what an agent promised in messages against the orders it actually submitted.

Key highlights

• Supports multiple LLM providers (OpenAI, Claude, Gemini, DeepSeek, OpenRouter) and can assign a different model to each power
• BFS pathfinding on the game map surfaces nearest threats and opportunities for context-aware order generation
• Resume interrupted games or branch into “critical state analysis” runs from any phase
• Batch experiment runner spins up parallel games with per-model token limits and custom prompt directories
• Power-specific system prompts and instruction templates shape each agent’s diplomatic personality

Verdict

Worth a look if you’re benchmarking multi-agent negotiation, studying alliance dynamics, or just want to watch Claude backstab GPT-4o over the Balkans. Skip it if you need a lightweight LLM wrapper—this is a full game engine with heavy prompt orchestration.