Tech • IA • Crypto
Agent Battle: Mine the most diamonds in 45 minutes
4/10TL;DR
A technical workshop pits artificial intelligence agents against each other in an optimization race to extract as many diamonds as possible in Minecraft within a limited time.
KEY POINTS
A real-time agent competition
Participants were invited to design and fine-tune agents capable of mining diamonds within 35 minutes, with each attempt lasting 5 minutes. The goal was simple on the surface: obtain the highest number of diamonds, with a live leaderboard adding competitive pressure.
Three major technical objectives
The exercise aimed to demonstrate the creation of a managed agent, the impact of configuration choices, and the importance of continuous improvement through evaluations. Participants had to adjust key parameters such as the system prompt, the model used, and integrated tools to influence agent behavior.
A standardized environment for performance comparison
All competitors started with the same seed, the same initial equipment, and an identical environment, ensuring fair comparison. Each agent operated in a Minecraft clone connected to a Mind Flare bot, without a direct visual interface, but with structured commands such as moving or interacting with blocks.
Optimization levers focused on code and strategy
Participants mainly worked in a my_agent.py file, where they could modify the model, write a system prompt, and integrate specific skills. The use of MCP servers and preconfigured modules helped accelerate testing while still allowing room for experimentation.
The importance of rapid evaluations
A fast evaluation system of about one minute enabled efficient iteration. This approach, described as “hill climbing,” involves continuously testing, measuring, and adjusting to improve agent performance within a limited timeframe.
A constraint on token efficiency
Beyond the number of diamonds, a decisive criterion was the diamonds/tokens ratio. In case of a tie, the agent that used fewer computational resources was favored, pushing participants to prioritize fine optimization rather than relying on heavier models.
Close results and unexpected performances
Near the end, several participants were tied, with an observed ceiling around 19 diamonds. One competitor ultimately surpassed this score in the final minutes, illustrating the impact of late-stage optimization. However, an anomaly was noted with a score showing 0 tokens, calling some intermediate rankings into question.
Technical constraints in real-world conditions
Connection issues related to Wi-Fi disrupted some participants, particularly involving Cloudflare, highlighting the practical challenges of deploying agents in shared, high-load environments.
CONCLUSION
This competition highlights the rise of autonomous agents and the importance of fine-tuned configuration, where raw performance and resource efficiency become inseparable.