The transition from human-operated Software-as-a-Service (SaaS) to autonomous, stateful execution (Service-as-Software, or SaaSw) is blocked by a techno-economic wall. Deploying monolithic Large Language Models (LLMs) as continuous execution engines yields prohibitive marginal costs, while naive "agentic" routing fails catastrophically under domain drift because it assumes ground-truth observability at inference. We introduce the Autonomous Service Fabric (ASF), an ML-systems architecture for constrained, calibrated routing under unobservable correctness. The ASF shifts stateful execution from monolithic reasoning to a dynamically calibrated probabilistic router governed by Adaptive Conformal Inference (ACI). We concretize the execution environment as a Partially Observable Markov Decision Process (POMDP), defining explicit, computationally trivial Bayesian belief updates over transaction stages to inform routing thresholds. To guarantee enterprise safety, externally committed actions are validated against a Canonical Ledger to mitigate structural error, while irreversible actions are structurally gated via Role-Based Access Control (RBAC). Validated on the World of Workflows (WoW) benchmark, the ASF dynamically adapts to distribution shifts and queue backpressure, achieving a 5–20× reduction in cost per successful outcome relative to LLM-only baselines and state-of-the-art routing baselines (RACER, RouteLLM) while mathematically bounding catastrophic semantic errors.
Anteneh Tessema (Sat,) studied this question.