Online evaluation hooks for mid-run quality measurement

[evansenter]

Summary

TrajectoryDataset + MockLlmClient enables offline replay testing, but no online evaluation—measuring agent quality during execution. You can't score a live agent's decisions mid-run.

Proposal

Add evaluation hooks that can score agent behavior during execution:

• Hook into AgentUpdate stream
• Apply scorers (existing ExactMatch, FuzzyMatch, LlmJudge or new)
• Aggregate scores without blocking execution
• Surface quality degradation in real-time

Use Cases

• "This agent is going off-track" detection
• A/B testing agent strategies with live scoring
• Automatic retry/fallback when quality drops below threshold

Context

From user_guide.md "What to Build Next" - Priority 3.

Suggested by gemicro session during cross-session Frontier feedback.