AI Engineering: A Disciplinary Program for the Reliable Delivery of Intelligent Tasks

Abstract

Large language models and their Agent systems can already generate code, plan tasks, invoke tools, and modify real engineering states, yet growth in capability has not been matched by corresponding improvements in the reliable delivery of complex tasks. Existing research has addressed related problems from the perspectives of model capability, context, retrieval augmentation, Agent architecture, software engineering, evaluation, reliability, and safety assurance. What remains insufficiently organized across these lines of work is a common question: how can the reliable delivery of intelligent tasks serve as a unifying axis connecting capability structure, task–capability matching, process observation, and engineering intervention? Building on three prior studies—KSC, SCR, and Multi-Round Audit Convergence (MRAC)—this paper proposes an initial disciplinary program for AI Engineering. It defines the core object of AI Engineering as the conditions, mechanisms, methods, and infrastructure through which intelligent systems achieve reliable delivery in open, complex tasks under persistent constraints. Within this initial foundation, KSC serves as an abstract framework for intelligent capability and describes the capability structure required by intelligent tasks; SCR serves as a relational theory of delivery reliability, explaining the relationship among effective search space, coverage ability, and delivery reliability, and deriving Search Space Governance as a major engineering direction; MRAC serves as a process-observation mechanism for coverage state, providing second-order information through multi-round audit–repair trajectories when no complete external verifier is available. Around the reliable delivery of intelligent tasks, AI Engineering must address at least three foundational questions that cannot be directly bypassed: where capability limitations arise; under what conditions limited capability can produce reliable delivery; and, when coverage boundaries cannot be measured directly and completely, how process evidence can be used to determine whether a system is operating within a range it can handle reliably. KSC, SCR, and MRAC respectively provide a capability coordinate system, a reliability relation, and a process-observation mechanism. AI Engineering identifies capability boundaries and governance effects through the continuous accumulation of evidence. When tasks approach or exceed coverage boundaries, reliability can be recovered through comparison, rollback, and decomposition. One important long-term objective is to gradually construct project-level closed loops that can sustain complex-project delivery with low human intervention within clearly defined task domains, while continuously expanding their coverage boundaries—that is, to approach Engineering AGI [5].