Simcity_Builtit_Agent
An AI agent that learns, plays, and gets better at SimCity BuildIt with every move.
Core trick
Agent learns from long gameplay videos, then plays by reading screen pixels and recovers from tutorial blockers.
Description
An AI agent that learns, plays, and gets better at SimCity BuildIt with every move.
README excerpt
# SimCity BuildIt Autonomous Agent — Powered by Gemini 4 Argon An autonomous closed-loop game agent built from scratch during the hackathon using **Gemini 4 Argon (`gemini-4-argon`)** that **Learns**, **Plays**, **Reasons**, and **Self-Heals**.  --- ## 🎯 Problem Statement Open-world 3D city-building games like **SimCity BuildIt** break traditional static bots due to shifting isometric cameras, floating 3D UI bubbles, multi-step material dependencies, and unexpected tutorial locks. Our goal is simple: **Build an autonomous agent that can Learn from long human gameplay videos, Play live via vision-grounded control, and Self-Heal when unexpected blockers occur.** > **Built at the Hackathon:** Both `learning_from_long_videos` and `executing_agent` were built from scratch during the hackathon using `gemini-4-argon` (`thinking_level="high"`) immediately after the problem statements were announced. --- ## 🏆 How We Address Both Hackathon Problem Statements | Problem Statement | How Our Project Implements It | Verified Telemetry | | :--- | :--- | :--- | | **1. Million Token Engineering** | `learning_from_long_videos/long_video_agent.py` processes a **44.5-minute YouTube gameplay video (201,094 tokens)** in a **single pass** without chunking drift, generating **40 consistent JSON skill files** (`18,974` continuous output tokens). During execution, all 40 JSON files are loaded into context simultaneously (`~21K` text tokens/call)