Subagent bursts
Also known as sub-agent bursts, bursty agent traffic
In plain English
An agent can launch several short-lived helpers together, suddenly adding requests and new context for the server to hold.
Technical definition
Subagent bursts are short periods of increased inference demand caused by an agent launching multiple subordinate tasks with their own request sequences.
Engineering details
A parent session can have a long reusable prefix while its new branches start with fresh context. These branches create overlapping prefill and decode work and temporarily enlarge the KV-cache working set. Their start times and dependencies matter as much as their request count.
Why it matters
A high average cache-hit rate can hide periods of heavy new-prefill demand. Capacity planning must account for burst timing, cache eviction, and the latency of branches whose completion blocks the parent task.
How to read it in InferenceX
The Rubin article names subagent bursts alongside multi-turn sessions, long context, and high prefix reuse. AgentX comparisons therefore concern a time-varying request pattern, not a constant batch of identical prompts.
Source material
See the concept in real benchmarks
Rubin NVL72 Agentic Inference: 67x better Performance per Dollar
Jensen Sandbagging Performance Again, 2x more Annual Profit Per GigaWatt, The More you Buy, The More you Earn, AgentX, InferenceX, Extreme Co-Design
A Brief Overview of Agentic Workloads
Multi-turn sessions, long contexts, and near-total prefix reuse make agentic inference a systems problem, and change what a benchmark has to measure