QueueBreak: A Trace-to-Diagnosis Pipeline for Tail Latency in Agentic LLM Services

Agentic large language model (LLM) services execute a user request as a workflow that interleaves model invocations with external tool calls. When workflow p99 grows, end-to-end latency alone cannot reveal whether the tail accumulates before execution starts or within execution; these regimes require different operational remedies. We present QueueBreak, a workflow-aware trace-to-diagnosis (T2D) framework that reconstructs request arrival, first-step start, and completion from lightweight workflow logs and step traces. QueueBreak decomposes each workflow before aggregating tail statistics, reports a Queue Dominance Signature, and checks the diagnosis through load-reduction intervention. Across load escalation, dependency degradation, and intervention, the evaluated service shows the same pattern: tail growth is dominated by pre-execution waiting while execution remains stable. In a mixed-workload sweep, total p99 rises from 21.25s to 165.40s as pre-execution p99 rises from 11.67s to 144.04s, whereas stress execution p99 is 16.80s. With tool failure $p_fail=0.2$, total p99 reaches 978.81s and the mean preexecution fraction reaches 0.923. Reducing load to one quarter of the stress rate cuts pre-execution p99 from 579.2s to 17.3s while execution p99 changes only from 18.0s to 17.5s. These results identify admission control and workflow-aware capacity planning as primary QoS recovery levers.