新兴代理型AI软件基础设施市场：从SaaS到代理型AI

From SaaS to agentic AI Digital and Analytics/ArticleMarch 10, 2026 Between the late 1990s and the mid-2010s, the rise of software-as-a-service (SaaS) fundamentallydismantled the on-premises application market. Early SaaS pioneers such as Salesforce (founded in 1999)demonstrated that multi-tenant, subscription-based delivery could outperform licensed, install-and-maintain software on cost, speed, and scalability. By the late 2000s, SaaS adoption accelerated rapidly ascloud infrastructure matured, culminating in the 2010–2015 period when SaaS became the dominantmodel for new enterprise software purchases. These SaaS companies built their new business on the backof cloud software infrastructure. These foundational capabilities, spanning cloud infrastructure, cloudplatform services, cloud operations, and cloud financial and commercial management, were crucialprerequisites for SaaS adoption at scale and had to mature before the SaaS market could fully take off. Much like the industry’s transition from on-premises software to SaaS, agentic AI will also usher in a newsoftware delivery model, leading to very different underlying software infrastructure (see figure 1). Asoftware infrastructure model is beginning to take shape for agentic AI, providing the foundationalservices required to enable the transition from application-centric SaaS platforms to agent-centric agent-as-a service (AaaS) systems. In the following sections we describe how this agentic AI softwareinfrastructure landscape is shaping up and then offer some investment hypotheses to capitalize on thisshift. The evolving agentic AI software infrastructurelandscape The agentic AI software infrastructure market remains nascent and rapidly evolving. While definitive,mature platforms akin to today’s cloud databases or container orchestration layers do not yet exist, clearpatterns of capabilities are crystalizing. These early patterns provide a valuable lens into how investors,vendors, builders, and enterprise adopters should think about future infrastructure investments. At its core, agentic AI infrastructure must support a set of emerging capabilities that extend well beyondtraditional AI platforms. First, solutions must be capable of running in secure, standalone environmentsthat meet enterprise compliance and data governance requirements. Second, they must support dynamicdiscovery and integration with external services and data sources, enabling agents to act on up-to-datebusiness context. Closely related is the ability for agents to discover and collaborate with one another,sharing work and knowledge rather than operating as isolated silos. Persistent memory, both short-termand long-term, is essential, as is the ability to store evolving decision context that reflects how agentslearn over time rather than simply recording transaction results. From an operational perspective,infrastructure must enable secure, cost-effective scaling of agents at enterprise scale, paired with toolingto measure and understand the economics of agent execution—a prerequisite for cost-effectiveoperations, pricing, and commercial viability. Finally, the infrastructure must support the integration ofthird-party agents and services, enabling ecosystems of agents to coalesce around value streams. The emerging landscape of agentic AI infrastructure can be logically decomposed into a series offunctional layers (see figure 2): •Agentic platforms, including anthat provides the core execution layer whereagentic AI runtimereasoning, goal management, dependency resolution, short-term memory, and behavioral logic reside.This is the locus of autonomous decision-making. Complementing this is anagentic AI orchestrationlayer, responsible for coordinating agents; managing scheduling, logging, and security; reconcilingcontext and memory; and enabling agents to discover one another and external services. •MCP (Model Context Protocol) runtime and servers,which allow agents to interact with externalservices and standardized APIs, acting as gateways and registries for MCP servers that agents access.1•Agentic context stores, including context and long-term memory infrastructure that persistsknowledge and evolving agent state across interactions and operational cycles, in addition toknowledgethat house domain content, ontologies, structured business context, and foundational data usedstoresby agents for informed reasoning. •Agentic market capabilities, including anto make agents readily discoverable,agent registryagentto allow users to browse and manage available agents, andthat acceleratecatalogsthird-party agentsthe development of multi-agent systems. •Operations, governance, and security managementtooling that provides observability, healthmonitoring, life cycle controls, oversight, and security management for production-ready agents.•Financial and commercial managementlayers that enable cost allocation, showback/chargeback,and pricing and monetization models to be applied to agent ope