Back to feed
arXiv cs.CL·

Scaling Enterprise Agent Routing: Degradation, Diagnosis, and Recovery

Signal
78
Hype
15
In three linesStudy on agent routing at scale: with 110 agents and 584 tools, F1 accuracy drops 16–23 percentage points on under-specified requests. Analysis decomposes degradation into retrieval gap and confusion gap (10pp oracle ceiling loss). Embedding-based shortlisting recovers +10–11pp F1 at full scale across models.
Read source
Your take?
AI AgentsMulti-agentEvalsBenchmarks

Summary generated by Claude — human-verified