Summary

Five coordination metric data models exist (CoordinationEfficiency, CoordinationOverhead, ErrorAmplification, MessageDensity, RedundancyRate), but the runtime collection pipeline that computes these metrics from live multi-agent execution does not exist.

Design Spec Reference

• §10.5 M4: Coordination Metrics Suite

Scope

• Runtime collection for all 5 metrics (opt-in via config)
• Single-agent baseline data collection for error amplification (Ae)
• Embedding computation for redundancy rate (R) — technology TBD
• Integration with observability events
• Alert thresholds for orchestration overhead

Research Context

Scaling Agent Systems (arXiv:2512.08296) — 5-metric coordination suite:

• Ec (efficiency ratio), Ae (agent effectiveness), O% (overhead percentage), c (message density), R (redundancy)
• R-squared=0.524 predictive power for topology selection
• Tiered overhead thresholds: optimal 200-300%, under-coordination <100%, over-coordination >400%

MADQA (arXiv:2603.12180) — Per-agent effort efficiency:

• accuracy / normalized_turns_used metric
• Measures not just "did it work" but "how efficiently"
• Raw data already available in TurnRecord

See research/scaling-agent-systems.md and research/stagnation-detector.md in project memory.