Intent-Driven Fleets: An Agentic AI Framework for Cloud Elasticity

Abstract

The evolution of cloud infrastructure toward hyper-scale deployments has exposed the fundamental inadequacy of reactive, threshold-based auto-scaling mechanisms. As digital services grow to serve global user bases during concentrated seasonal demand windows, the gap between high-level business objectives and low-level infrastructure execution has widened into a structural operational failure. Intent-Driven Fleets (IDF) address this gap through an autonomous orchestration framework that coordinates specialized AI agents, Commander, Forecasting, Provisioner, and Efficiency via the Model Context Protocol (MCP), enabling infrastructure to reason about goals rather than execute pre-written rules. The framework proposes the Plan-Execute-Observe-Reflect (PEOR) cycle, a formalized iterative process in which infrastructure anticipates demand, breaks business intent down into dependency-based execution plans, accepts business-layer telemetry to provide a context in which decisions are made, and continually optimizes provisioning behaviour via long-term memory. Security is also achieved by a deterministic guardrail layer, which is directly implemented as part of the MCP server that ensures that agent actions are not unlimited financially by permitting only signed authorization tokens, which are verified prior to each tool call. Individually identified engineering issues of this architecture are: context window congestion when receiving full telemetry, tool-call latency buildup amidst multi-region provisioning sequences, and concurrency conflicts necessitating distributed intent locking. The IDF framework establishes intent as the most effective abstraction for managing hyper-scale cloud environments, pointing toward a genuinely autonomous operational paradigm where global infrastructure responds directly to business goals without requiring continuous human translation at every execution step.

DOI: https://doi.org/10.17762/ijisae.v14i1s.8220