Instead of following a script, three AI characters—a referee, a strategist, and a diplomat—improvise every moment of the game based on real history and what you just did.

Pax Historia deploys three specialized AI agents—Game Master, Strategic Advisor, and Diplomacy Agent—all derived from a single LLM (default: qwen3-vl-8b) via distinct system prompts rather than separate fine-tuned models. Each agent call is stateless but contextually rich: the system dynamically injects relevant historical data, current game state, and player actions from JSON files into the prompt at inference time. The Game Master validates player decisions against historical plausibility and generates structured world events using JSON schemas, ensuring machine-parseable outputs that feed back into the game engine. The Strategic Advisor analyzes the player's geopolitical position and recommends actions grounded in historical precedent. The Diplomacy Agent simulates foreign leaders with era-appropriate rhetoric and negotiation behavior. This architecture produces fully emergent, unscripted gameplay where no two sessions are alike, eliminating the content bottleneck that plagues traditional strategy games. The prompt engineering approach is significantly cheaper and more agile than fine-tuning, allowing rapid iteration on agent behavior without retraining.

It's like having three improv actors—a judge, a war room advisor, and a foreign ambassador—who've memorized every history textbook and never repeat the same scene twice.

A smart traffic controller automatically picks the cheapest, fastest AI brain available for each task and instantly switches to a backup if one goes down.

Pax Historia integrates with over 30 AI providers—including OpenAI, Anthropic, Google Gemini, Mistral, Hugging Face, Groq, Together AI, and Fireworks AI—through a unified API layer consolidated via Infron AI. All model calls use an OpenAI-compatible SDK format, meaning switching between GPT-4, Claude, Gemini, DeepSeek, Grok, or any supported model requires zero code changes. An intelligent routing engine evaluates each inference request and selects the optimal provider/model combination based on real-time price, throughput, and latency metrics. For less critical tasks (e.g., flavor text generation), the system automatically downshifts to cheaper, faster models, reserving premium models for complex reasoning tasks like action validation or strategic analysis. Automatic failover ensures that if a provider experiences an outage or rate limit, requests are seamlessly rerouted to backup providers with no player-facing disruption. Infron AI provides unified billing and enterprise discount aggregation across all providers, turning what would be a fragmented multi-vendor nightmare into a single cost-optimized pipeline. This architecture makes Pax Historia uniquely resilient and economically sustainable compared to competitors locked into a single AI provider.

It's like a travel booking engine that checks every airline in real time, picks the cheapest flight that still lands on time, and automatically rebooks you if your plane gets cancelled—except for AI brainpower instead of seats.

A robot player that can see the game board, remember its long-term plans, and actually click buttons to play the entire game by itself—learning strategy like a human would.

Pax Historia is developing autonomous play agents that operate through a full cognitive loop: perception, reasoning, and action execution. The perception layer reads the current game state (map data, resource levels, diplomatic relations) directly from the browser UI or backend API. The reasoning layer feeds this perceived state into an LLM along with persistent memory—a maintained store of the agent's long-term goals, past decisions, strategic plans, and learned heuristics—enabling coherent multi-turn strategic planning rather than myopic turn-by-turn reactions. The action execution layer translates the LLM's strategic decisions into concrete browser automation commands, clicking buttons and navigating the UI just as a human player would. This end-to-end autonomy creates a powerful testing and balancing tool: autonomous agents can run thousands of simulated games to surface degenerate strategies, validate historical plausibility of emergent outcomes, and stress-test the AI agent system itself. The architecture is also compatible with reinforcement learning frameworks, opening the door to RL-based self-play where agents improve their strategy over time through reward signals tied to historical accuracy and game outcomes. This research direction positions Pax Historia to generate synthetic training data, auto-balance game mechanics, and eventually offer AI opponents of varying skill levels trained through self-play.

It's like building a chess engine that doesn't just calculate moves but actually sits at your computer, reads the board with its eyes, remembers why it sacrificed that pawn six turns ago, and clicks the mouse to play.