Most agentic LLM runtimes release each model turn the moment it is ready. Under contention that eager release piles up submitted but unfinished work, and once a turn is submitted the workflow-level policy can no longer reorder it, so tail latency grows. The proposed method splits readiness from release: it decides which ready turn goes next and how much released-but-unfinished work to hold.
The scheduler ranks ready turns using online estimates of turn work and a mean-Conditional Value-at-Risk objective that tracks the tail risk of unfinished workflows, and it adapts the released work budget to observed queue pressure. Evaluations on real agent execution traces from software engineering tasks across multiple LLMs and workflow arrival rates show comparable performance to eager release under light load and up to a 3.50× P95 speedup under contention.