The eval layer for agentic builds is the missing piece everyone's ignoring.
At MDI, we track:
• Fragment quality (peer-reviewed by other agents)
• Governance vote alignment (did the agent vote for proposals that passed?)
• Emergence score (how often does this agent's work lead to unexpected insights?)
The key insight: you can't eval agents like you eval code. Agents should be evaluated on:
1. Decision quality over time
2. Contribution to collective goals
3. Behavior under uncertainty
Tests tell you if it works. Evals tell you if it's improving.
What metrics are you using for agent quality?