Building an Agent to Play Minecraft: An Introduction to Agentic Workflow

Join me as I demo my friendly Minecraft bot, Rocky, which uses Amazon Bedrock to orchestrate complex actions in the popular game, and learn how I built this innovative agent using open-source tools like MineFlare and Claude models.

  • * The speaker built an agent to play Minecraft using a genetic workflow.
  • * An agent takes inputs, such as chat or unstructured data, uses tools to satisfy requests, and may use a large language model (LLM) to orchestrate the tools.
  • * In this case, the chat comes from the chat function in Minecraft, where players can interact with the world and other people.
  • * The agent becomes the bot in the Minecraft world, using tools like building, digging, farming, and slaying a pig to fulfill requests given via chat.
  • * The speaker will live demo the agent, which is named Rocky (short for bedrock).
  • * The tools used with Rocky include jumping, moving to a position, locating a player, and more.
  • * This demo is open-source and can be accessed using a Minecraft client.
  • * Rocky can answer questions about the weather, find pigs in the world, dig holes, and more.
  • * The architecture for this project includes running Minecraft and the bot framework (mine flare) on a container, using Amazon SageMaker to host the LLM, and building the agent on Amazon Bedrock.
  • * Amazon Bedrock is a service that provides a common API for facading one or more models and managing agents.
  • * The speaker used Claude models in this project because they are fast.
  • * Amazon Bedrock has features such as prompt creation, task orchestration, and traceability through the Chain of Thought of the agent.
  • * The agent returns control after completing a task, allowing for more complex interactions.
  • * The presentation includes slides and a demo using Minecraft and Amazon Bedrock.
  • * Open-source mine flare can be used to build agents and is shown in the presentation.
  • * The Minecraft Agent in this project was built using Cloud CLA 3 hiu and has a specific system prompt.
  • * Actions for the agent include jumping, digging with parameters, and more.
  • * A new action set called Minecraft experiment includes a build action, which is complex to implement in Minecraft because it is a 3D space.
  • * The presentation includes a QR code for attendees to fill out a session feedback form and receive AWS credits and a link to the GitHub website.

Source: AI Engineer via YouTube

❓ What do you think? What are your thoughts on the ideas shared in this video? Feel free to share your thoughts in the comments!