research(routing): Agent Stability Index for behavioral drift detection #1841
Description
Research Finding
Paper: Agent Drift — Behavioral Degradation Quantification and Drift-Aware Routing (arXiv:2601.04170, 2025)
Proposes "Agent Stability Index" (ASI) measuring behavioral consistency over extended interactions. Quantifies how multi-agent systems degrade in persona, task coherence, and output style over long sessions. Drift-aware routing uses ASI as a routing signal alongside latency/quality metrics.
Applicability to Zeph
Known blind spot: Zeph has no runtime behavioral consistency metric. The Thompson Sampling router uses latency EMA + success/failure counts but has no signal for response coherence drift.
Two concrete improvements:
1. ASI as routing signal
Add a coherence_score to ProviderStats in the Thompson Sampling router:
- Measure response embedding similarity to recent N responses (sliding window)
- High variance → low coherence → penalize in routing prior
- Works with existing
EmbeddingStore— embed each response, compare cosine to window mean
2. Episodic consolidation to prevent context drift
When ASI drops below threshold during a long session, trigger a lightweight episodic consolidation (summarize recent N turns into a stable anchor) before the next LLM call. Prevents the gradual persona/context drift observed in sessions with many compactions.
References
- arXiv:2601.04170
- Zeph crates:
zeph-llm(router/thompson.rs,ema.rs),zeph-memory(semantic/)