Summary
In this live workshop hosted by Anthropic’s Applied AI team, engineers Ben and Jeff guide participants through building and deploying a managed agent using Anthropic’s hosted infrastructure. The challenge is deceptively fun — agents compete to mine the most diamonds in Minecraft within 45 minutes — but the real curriculum is serious: understanding how agent configuration choices (system prompt, model selection, MCP tools, and packaged skills) directly shape agent behavior and output quality.
Participants work with a Python agent harness (my_agent.py) connected to a Minecraft clone via a Mineflayer bot, with a suite of MCP tools for navigation and block interaction. Crucially, the competition is scored not just on raw diamonds collected but on diamond-to-token ratio, forcing participants to write lean, precise system prompts rather than defaulting to the heaviest available model.
The workshop’s third pillar is eval-driven iteration — the same hill-climbing methodology Anthropic uses internally to improve production agents. With a fast one-minute eval loop, participants can rapidly test prompt changes and compare results on a live leaderboard. It’s a compact, hands-on introduction to the full agent development lifecycle: configure, evaluate, iterate.
📺 Source: Claude · Published May 23, 2026
🏷️ Format: Hands On Build
