AI Agents Defeated Minecraft Dragon in Nine Minutes
A new pairing of models successfully navigated complex gaming tasks in under ten minutes for less than one dollar.
Updated on Sept. 21, 2026 in Artificial Intelligence
Live Poll
Do you believe using specialized AI tag teams is the most efficient way to advance technology?
AI agents Jev and Astra completed a Minecraft Ender Dragon defeat in 8 minutes and 43 seconds. This performance follows a previous unsuccessful 141-hour attempt by the Astra model, demonstrating a marked shift in task-specific agent efficiency.
Why it matters
The successful integration of System 1 models like Jev enables faster, resource-efficient decision-making for repetitive tasks compared to traditional System 2 large language models. This approach significantly lowers the compute cost for complex agent-based workflows.
The operation cost less than 1 dollar in tokens for the full run. Jev and Astra achieved this by offloading immediate character movements to a specialized model while using the other for planning.
The players
TypeSafe AI
A developer of specialized AI models focused on System 1 architecture and efficient decision-making.
Diogo Almeida
The founder of TypeSafe AI who leads development on efficient agentic model architectures.
Ronak Malde
The researcher who performed the Minecraft technical test to evaluate the agent's performance.
The details
The integration utilizes Astra as a high-level planner that determines objectives and waypoints for the agent. Jev then executes the specific in-game actions using WASD, mouse, and click controls. By receiving unstructured state information and selecting from predefined options rather than generating free-form text, Jev bypasses the latency of standard generative models.
Timeline
September 2026: The AI agents defeated the Minecraft Ender Dragon.
September 2026: The previously recorded 141-hour failed Minecraft run gained significant attention.
The Tech Race
This performance marks a significant departure from previous Minecraft-based AI research that often relied on massive, multi-day compute cycles to achieve objectives. The project effectively redefines the efficiency standard for agentic tasks by prioritizing specialized, rapid-execution models.
Developers can access the full code for this integration on GitHub to implement or test these agentic workflows in their own environments. The low cost of operation makes this model configuration highly accessible for hobbyists and researchers interested in rapid-action AI.
The takeaway
The success of Jev and Astra proves that System 1 models can handle complex, multi-step tasks at a fraction of the cost previously deemed necessary. Watch for new benchmarks on GitHub that test these models against non-gaming environments to see if this speed holds under real-world constraints.
Further reading
For broader context on how agentic architectures are evolving, visit the Artificial Intelligence section.
Source note: This article includes information reported by MakeUseOf.
Live Poll
Do you believe using specialized AI tag teams is the most efficient way to advance technology?
