Agentic AI systems powered by large language models run as autonomous multi‑step pipelines that can invoke tools on the fly, store observations in memory, and iterate until a task finishes or a resource budget is exhausted. Unlike classical workflows with a fixed graph, the total compute, bandwidth and end‑to‑end latency of such pipelines are unknown at dispatch; they emerge at runtime from intermediate observations and stochastic token sampling, yielding heavy‑tailed, multimodal random variables. This unpredictability creates a dilemma for resource orchestrators: under‑provisioning leads to throttling or missed deadlines, while over‑provisioning wastes scarce edge or cloud capacity. Existing schedulers assume a fully known pipeline graph, and edge resource‑management approaches rely on parametric or stationary demand models, neither of which fits the unknown, non‑stationary nature of agentic pipelines. We formalize this as the uncertainty‑aware agentic orchestration problem, whose goal is to reserve compute and bandwidth at dispatch that cover the actual pipeline demand and to verify deadline feasibility with a prescribed probability. To address it we propose Conformal Reservation for Orchestration (CRO), an online algorithm that converts a sliding window of past execution traces into dispatch‑time allocations and provides finite‑sample, distribution‑free coverage guarantees via conformal prediction. Simulations on heterogeneous task mixes with multimodal and heavy‑tailed profiles show that CRO is the only evaluated method that simultaneously meets target probabilistic coverage across all risk levels, and among coverage‑feasible methods it achieves the lowest over‑reservation, roughly halving the resource waste of the Chebyshev‑bound baseline. Review