The Seed – An LLM Agent Framework for Game Developers

中文版README

Overview

The Seed is an open-source framework designed for game developers. Its goals are:

• Allow games to expose their state / actions to an LLM-Agent through a unified integration protocol, enabling the LLM to observe the game, issue actions, and interact with players.

• Enable LLMs and traditional game AI to form a complementary workflow, handling decision-making and execution together under a controlled compute budget.

---

Design Principles

1. Give the Agent a clear and well-defined entry point

   • The basic assumption is that the game is willing to expose an interface layer of “observations / actions / events.”
   • Integration should be minimally invasive—more like attaching a module, not rewriting the game’s logic.
   • The framework aims to stay compatible with different engine architectures and allow light-weight, maintainable integration.

2. LLM for high-level decisions, game for execution

   • LLMs excel at: situational understanding, plan generation, strategy reasoning, and behavior explanation.
   • Game-native AI (behavior trees / state machines / rule systems) excel at: pathfinding, micro-actions, condition checks, and frame-level logic.
   • The Seed follows this separation of roles:
     LLM generates intentions → the game executes them.
     Achieve stable behavior with minimal model calls.

3. Not tied to any specific gameplay

   • Provides an extensible Action / Observation / Tool protocol.
   • Does not pre-define semantics such as “attack / gather / build”; each game defines its own actions and data structures.
   • The framework organizes these definitions into LLM-friendly prompts and tool interfaces so that different games can build their own Agent styles within a shared framework.

4. Iterate from real integration experience

   • The project is still under active development.
   • Priority: reduce the steps and code needed for a new game to “go from zero to having an Agent running.”
   • Use real projects to refine scaffolding, examples, debugging tools, and best practices.
   • The long-term goal is to provide a framework that is practical, well-documented, and easy to introduce to your team.

---

Current Project Status

• Stage: PoC / Early Prototype
  • ✅ Completed: overall architecture draft & agent interaction workflow
  • ✅ Completed: initial version of the game-side API (observation / action / event)
  • ⏳ In progress:
    • The Seed core Agent protocol
    • Demo Agent for OpenRA (e.g., auto-econ / auto-battle)
    • First version of README / documentation / integration guides

---

Roadmap

Phase 0 – OpenRA Validation (⏳ Ongoing, ~45–60 days)

• Define a basic RTS-oriented Agent protocol (observation / action / tick / event)
• Deliver an “OpenRA + Agent Demo” that works out-of-the-box
• Prepare developer documentation:
  • How to integrate The Seed into a game
  • How to write a minimal LLM-Agent controlling one faction

Phase 1 – Framework Stabilization & Documentation

• Extract a Core SDK decoupled from any specific game
• Improve:
  • Agent lifecycle management
  • Tick / planning / memory / logging
  • Adapters for cloud or local LLM models

Phase 2 – Multi-Game Integration & Community Growth

• Add a second supported game (priority: strategy / simulation)
• Build a Sample Integrations Collection
• Host developer-oriented activities:
  • Hackathon / Game Jam
  • Online workshops and technical sharing

Phase 3 – Agent Ecosystem

• Propose an Agent Description Standard to support:
  • Sharing strategies across different games
  • Community-created Agent roles
• Explore additional features:
  • Multi-Agent cooperation
  • Coach / spectator-type Agents
  • Replay analysis / explanation Agents

---

How to Contribute

• ⭐ Star the repo — follow updates and support the project
• 🐛 Open Issues — ideas, feedback, bug reports
• 🔧 Submit PRs — docs, improvements, examples
• 📣 Spread the word — share with game devs or AI enthusiasts

---

Getting Started

TODO

1. Environment Setup

TODO