Agent Gathering: Avatar Siege

Agent Gathering turns real AI-system behavior into card-game strategy.

In Avatar Siege, players build networks of Agents, Firewalls, and tools while treating an opponent's context window as a visible game resource. Prompt attacks add retained context, sub-agents traverse defensive layers, and Compaction can recover headroom before an Avatar reaches redline.

The default match puts a human player against GPT-5.6 Luna. Luna chooses only from legal actions, while a deterministic server referee validates and applies every move, so the model supplies strategy but never controls the rules. The interface makes attacks visible: it shows what was hit, whether the attack landed, and how each Avatar's retained context changed.

I used Codex as my primary product and engineering collaborator to turn the existing card-game foundation into this playable Build Week edition. Codex helped implement and verify the persistent Avatar runtime, deterministic rules, OpenAI adapter, attack feedback, replay, privacy boundaries, and browser tests. GPT-5.6 powers Luna's bounded strategic decisions and the live AI interactions that become game mechanics.

Built With

• codex
• css
• fastapi
• gpt-5.6
• html
• javascript
• openai-api
• playwright
• pytest
• python
• uvicorn