Abstract
As artificial intelligence (AI) systems evolve from stateless chatbots to autonomous multi-step agents, prompt engineering (PE), the discipline of crafting individual queries, proves necessary but insufficient. This paper introduces context engineering (CE) as a standalone discipline concerned with des igning, structuring, and managing the entire informational environment in which an AI agent makes decisions. Drawing on vendor architectures (Google ADK [Agent Development Kit], Anthropic, LangChain), current academic work (ACE framework, Google DeepMind's intelligent delegation), enterprise research data (Deloitte, 2026; KPMG, 2026), and the author's own experience building a multi-agent system (MAS), the paper proposes five production -grade context quality criteria: relevance, sufficiency, isolation, economy, and provenance, and frames context as the agent's operating system. On this foundation, the paper introduces two higher-order disciplines. Intent engineering (IE) encodes organizational goals, values, and trade-off hierarchies into agent infrastructure, ensuring that well-contextualized agents pursue the right outcomes. Specification engineering (SE) extends this approach to creating a machine -readable corpus of corporate policies, quality standards, organizational agreements, and instructions that ena bles autonomous and coherent operation of multi-agent systems at scale. Together with prompt engineering and context engineering, these disciplines form the cumulative four-level pyramid maturity model of agent engineering (hereafter, the Pyramid), in which each level subsumes the previous one as a necessary foundation rather than superseding it. Independent convergence of several authors toward the same taxonomy in early 2026 confirms that the identified structure reflects the actual composition of the field. Enterprise data reveals a gap: while 75% of enterprises plan to deploy agentic AI within two years (Deloitte, 2026), agent deployment has already surged and retreated as organizations confront the complexity of scaling (KPMG, 2026). The Klarna case illustrates a dual deficit, contextual and intentional. The conclusion: whoever controls the agent's context controls its behavior; whoever controls its intent controls its strategy; whoever controls its specifications controls its scale. Since the launch of ChatGPT in late 2022, the interface between organizations and large language models (LLMs) has quietly but decisively changed shape. Users who once typed single questions into a chat window now deploy autonomous agents that call APIs, browse the web, write and execute code, and run for hours without human intervention. Composing effective prompts, the craft called prompt engineering, remains indispensable when the task fits within one requestresponse cycle; it was never designed, however, for the architectural pressures that surface when an LLM becomes a component inside an orchestrated multi-step system. An agent that plans a twenty-step workflow, delegates subtasks, and commits real-world actions operates under constraints that no single prompt can anticipate. That structural gap motivates the present paper. The response to this gap has been emerging from multiple directions simultaneously. LangChain published a programmatic definition of context engineering in early 2025; Anthropic incorporated the concept into its developer documentation; Google's Agent Development Kit operationalized context pipelines at the platform level. Academic work followed: the ACE framework formalized context layering, and Tomašev et al. (2026) proposed an intelligent delegation model that treats context boundaries as first-class architectural decisions. Enterpris e surveys confirm that agent deployment is outpacing governance. According to Deloitte (2026; N = 3,235, 24 countries), approximately 75% of organizations plan agentic AI deployment within two years, yet only 34% report using AI to deeply transform their business. KPMG's quarterly tracking (2026; N = 130, US Csuite) captures the velocity: agent deployment surged from 11% in Q1 to 42% in Q3 2025, before pulling back to 26% in Q4 as leaders shifted from pilots to production-grade systems, with average annual AI budgets reaching $124 million. Here we synthesize these converging threads, ground them in the author's three-year practice of working with numerous LLMs and building a multi-agent compliance system, and propose a unified framework with five production-grade quality criteria and a diagnostic taxonomy for context degradation. Yet even when an agent has access to all the right data, a critical question remains unanswered: to what end? An agent that sees everything it needs may still optimize for the wrong objective. The Klarna case (Section 12) illustrates how a technically capable agent can erode brand equity when corporate intent is left implicit. This observation motivates the paper's broader claim: context engineering is necessary but not terminal. Two additional layers, intent engineering and specification engineering, extend the logic from "what the agent sees" to "what the agent wants" and ultimately to "what the corporation demands at scale." The four disciplines together form the cumulative four-level pyramid maturity model of agent engineering (the Pyramid), whose structure is developed in detail in Section 17. This paper focuses on text-based LLM interactions and agentic architectures; the specific challenges introduced by multimodal inputs (images, audio, video) fall outside its scope, although the proposed framework applies to them in principle. The paper proceeds as follows. Sections 2-4 establish the baseline: prompt engineering as a craft, the architectural distinction between an LLM and an agent, and three tiers of agentic deployment. Sections 5-9 build the core argument for context engineering, covering its scope, architecture, and quality criteria complemented by Breunig's context-rot taxonomy. Sections 10-12 turn to economics, enterprise governance, and the Klarna cautionary case. Section 13 formalizes the transition from prompt craft to state engineering. Sections 14-16 introduce intent engineering, specification engineering, and agent memory as an infrastructure layer. Section 17 presents the Pyramid as a cumulative maturity model. Section 18 concludes.
Showing the abstract — retrieve the full paper via the Exa API.