Tomasz Tunguz Blog · 2026-01-22
· 133d
AI Managing AI: The Emergence of Specialist Agents and Orchestration Markets
As AI tool-calling accuracy has crossed 90% performance thresholds, frontier models are evolving into executive orchestrators that route tasks to specialized agents rather than handling everything monolithically. This shift is creating new startup opportunities in domain-specific AI agents that integrate into constellations managed by larger models, with economics favoring distilled, efficient specialists over trillion-parameter generalists.
Metrics in this report
Function-calling Accuracy (GPT-4, 2 years prior)
<50%
historical
GPT-4 model on function-calling tasks
Function-calling Accuracy (SOTA models, present)
>90%
median
State-of-the-art models on function-calling benchmarks
Inference Speed Improvement via Distillation
60%
relative
Distilled models vs original
Model Size Reduction via Distillation
40%
relative
Distilled models vs original
Performance Retention Post-Distillation
97%
median
Distilled models retaining original performance