CHAPTERS
- 0:20 – 0:51
Agent Battle kickoff: Diamond mining with managed agents
Ben and Jeff introduce the session as a timed “agent battle” where participants use AI agents to mine diamonds in Minecraft. The framing is both playful and practical: use agents (not manual play) to optimize performance under constraints.
- •Workshop format: competitive agent run in a Minecraft-like environment
- •Hosts introduce themselves and the Applied AI/Anthropic context
- •Goal is to mine diamonds via an agent rather than playing directly
- 0:51 – 1:51
Three learning goals: deploy, configure, and iteratively improve agents
Ben outlines the educational objectives behind the competition. Participants will learn to deploy a managed agent, see how configuration choices affect behavior, and use eval-driven iteration (“hill climbing”) to improve results.
- •Build and deploy a managed agent hosted on Anthropic infrastructure
- •Understand configuration levers: system prompt, model choice, skills/MCPs
- •Practice eval-driven iteration to improve agent behavior over time
- 1:51 – 3:07
Battle rules and scoring: timeboxing, submissions, and token efficiency
The competition constraints are clarified: a limited build/experiment window, 5-minute mining runs, and leaderboard tracking. Ties are broken by token efficiency, pushing participants toward lean prompts and efficient models rather than brute force.
- •~35 minutes to build and experiment; timer governs the session
- •One accepted run per person (top run counts), runs last 5 minutes
- •Leaderboard and chat provide real-time feedback and interaction
- •Tie-breaker: best diamonds-to-tokens ratio encourages efficiency
- 3:07 – 3:37
Harness overview: Mineflayer bot + MCP tool-based control (no visuals)
Jeff explains the technical harness powering the battle. Participants control a Mineflayer bot through shipped MCP tools (e.g., mining, movement), operating without a visual interface and focusing on tool-driven planning.
- •Minecraft clone connects to a Mineflayer bot
- •No reliance on visuals; control happens via tools and state
- •MCP tools include actions like mining blocks and moving near targets
- •Optimization focus is on agent behavior, not manual gameplay
- 3:37 – 4:07
Fair starting conditions: same seed and reset kit for everyone
The environment is standardized to keep the competition fair. Every reset uses the same world seed and the same starting kit, removing variance and emphasizing agent strategy and efficiency.
- •All participants start from the same world seed
- •Resets preserve identical seed and starting inventory/kit
- •Performance differences come from agent configuration and strategy
- 4:07 – 4:38
Where to modify the agent: repo layout and key files/knobs
Jeff points participants to the main customization surface in the repository. The central file is `my_agent.py`, where users can change the model, system prompt, skills, and optionally adjust MCP server behavior.
- •Work happens in the provided repo (e.g., `/agentbattle`)
- •Primary customization file: `my_agent.py`
- •Adjustable levers: model string, system prompt, shipped skill vs custom skill
- •Optional: modify or extend an included MCP server
- 4:38 – 5:08
Execution plan: iterate quickly with evals and ask for help
Participants are encouraged to try ideas, run evals, and iterate rapidly. Organizers will circulate to help debug or discuss strategies, maximizing productive experimentation during the limited time window.
- •Run evals frequently to test improvements quickly
- •Iterate on prompt/model/tooling to increase diamond yield
- •Organizers available for troubleshooting and strategy discussion
- 5:08 – 5:39
Countdown start: competition begins and leaderboard activity starts
The timer is reset and the battle begins. Jeff notes that participants are already submitting runs and results are appearing on the leaderboard in real time.
- •Official reset/refresh of the environment and start of the timer
- •Participants begin runs immediately
- •Leaderboard shows early results and participation momentum
- 5:39 – 6:09
Troubleshooting connectivity: CloudCode skill and network workarounds
Jeff provides practical help for setup issues, including a CloudCode skill in the repo and a suggested command to mitigate connectivity problems. Conference Wi‑Fi load is identified as a possible cause of Cloudflare-related issues.
- •CloudCode skill included to assist with setup and onboarding
- •Acknowledges Cloudflare/connectivity issues under conference Wi‑Fi load
- •Shares a command workaround that helped at least one attendee
- 6:09 – 7:12
Final minutes drama: ties, suspicious token counts, and new high score
As time runs low, Jeff calls out a three-way tie and an anomalous “zero tokens” entry affecting standings. A participant breaks the apparent ceiling (19 diamonds), adding suspense just before the timer ends.
- •Leaderboard tightening: multi-way tie near the end
- •Anomaly: top entry showing zero tokens prompts investigation
- •19 diamonds appears to be a top tier—then someone exceeds it late
- •Encourages a last run as the clock winds down
- 7:12 – 8:44
Time’s up and next steps: winners called to identify techniques
The competition ends and a clear winner is announced, though some placements are unclear due to token-count anomalies. Organizers invite top performers to come forward to resolve standings and share what worked.
- •Timer ends; winner announced
- •Second/third unclear due to suspicious zero-token entry
- •Top performers (notably those with 19 diamonds) asked to talk with hosts
- •Wrap-up and group acknowledgment
