--- title: "Why Today's AI Agents Don't Scale" author: "Daniel Gorld" author_role: "Consulting Director, cbs CX — The cbs Group Salesforce Consultancy" author_url: "https://cx-waves.com/about" publisher: "CX-Waves" canonical_url: "https://cx-waves.com/nodes/ai-agents-dos-phase" date_published: 2026-06-01 date_modified: 2026-09-20 language: en --- # Why Today's AI Agents Don't Scale Source: Daniel Gorld, CX-Waves — https://cx-waves.com/nodes/ai-agents-dos-phase (published 2026-06-01, updated 2026-09-20) ## Why are current AI agent systems difficult to scale, and what is needed to overcome this challenge? - **Lack of Orchestration Layer**: Current AI agent tools lack a unified orchestration layer for managing multiple agents simultaneously. This results in complex, manual processes for context handover, status tracking, and error handling when running several agents in parallel. - **Limited Visibility**: There is no visual interface where multiple agents can work side-by-side, making their operations and handovers traceable. Agents often function as "black boxes," obscuring their internal states and decision-making processes. - **Inconsistent Behavior**: Different agents and tools often exhibit inconsistent behaviors and interfaces, leading to varied conventions for communication and context exchange. This absence of standardization demands manual intervention for integration and interaction. - **Missing Enterprise Readiness**: Fundamental features required for enterprise-level multi-agent workflows, such as persistence, systematic handovers, comprehensive observability, and robust error handling, are not natively integrated into current tooling. This prevents scalable, reliable deployment in real-world business processes. Working with today's agent tools provides a familiar experience akin to MS-DOS in the early 1990s. This environment involves terminals, configuration files, command-line paths, and manual context manipulation between tools. While a single agent may be manageable, scaling to multiple parallel agents leads to a loss of overview and significant complexity. ### Diagnosis: We are in the DOS phase Current agent tooling, whether for coding, CRM configurations, or experimental multi-agent workflows, relies on command-line tools and declarative configurations like YAML or JSON. Data sources require manual connections, context windows are filled by hand, and outputs must be manually passed between agents. Managing one agent is feasible, but three agents create clutter, and ten agents, as in an end-to-end business process, cannot be cleanly orchestrated with existing tools. Consider a typical sales process where one agent qualifies leads, a second drafts quotes, and a third handles service initiation. Key challenges arise: determining which context is passed to whom, how one agent's decision is documented for the next, and how errors are identified and corrected. This mirrors the early 90s DOS experience, which was powerful for experts but not enterprise-ready at scale. ## What Windows brought back then The historical graphical revolution introduced three key advancements. First, a visual metaphor such as windows, desktops, and file managers provided a visible surface for running applications. Second, parallel visibility allowed multiple applications to operate simultaneously with clear status indicators. Third, consistent behavior across applications, including features like copy-paste, drag-and-drop, and shared interaction patterns, streamlined user experience. These three elements are largely absent in current AI agent systems. There is no unified visual surface where multiple agents can operate observably side-by-side, making their handovers transparent. Agents often run as opaque processes without a clear orchestration view or status trails. Furthermore, consistent behavior between agents is lacking; each may communicate differently, and context handover remains a largely manual task. Initial standardization efforts, like protocols for tool integration and early inter-agent communication approaches, are emerging. However, these are more akin to improvements in a DOS shell than a fundamental shift to a Windows-like operating model. A comprehensive operating model for enterprise multi-agent workflows—encompassing persistence, automated handovers, end-to-end observability, and automated error handling—is still undeveloped. ### Where the analogy breaks down It is important to acknowledge three distinctions where the historical analogy to DOS and Windows diverges, preventing a misleading comparison. Firstly, the vendor landscape changed. DOS was a single-vendor system replaced by Windows from the same vendor, whereas the agent world is diverse, with multiple models, providers, and architectures. A singular "Windows layer" is improbable; instead, a framework of open standards supporting several competing platforms is more likely. Secondly, user diversity has expanded. The transition from DOS aimed to elevate one type of user from command-line interfaces. Today, both power users and business process developers require solutions, presenting a more complex challenge than in 1995. Thirdly, the timeline for adoption is much shorter. The earlier transition took approximately a decade, while the current shift is projected for two to three years. This accelerated timeline means less opportunity for iterative refinement and emphasizes the need for rapid adaptation. ## The race for the orchestration layer Various influential groups are converging on the concept of an orchestration layer that manages handovers, status, error handling, and persistence for individual agents. LLM providers are enhancing their tooling to include orchestration capabilities, introducing standards for tool integration, desktop applications, and modular skill concepts. Hyperscaler platforms are developing their own agent studios, leveraging existing cloud infrastructure as a natural environment for multi-agent workflows. CRM and business application platforms contend that orchestration should reside where data, business logic, and processes are already established, namely within their ecosystems. Simultaneously, open frameworks and open-source projects are addressing the same challenge, offering vendor independence but often lacking enterprise readiness. The precise outcome of this race to develop the "Windows of the agent era" remains uncertain. It is likely that no single dominant layer will emerge, but rather a market with two or three leading orchestration platforms, unified by open standards. Significant investment from multiple strong actors suggests a dynamic and unsettled market. ## Consequences for B2B industrial companies Three key conclusions can inform B2B industrial companies navigating the current AI agent landscape, avoiding both excessive hype and undue caution. First, pilot projects are valuable as learning experiences during this "DOS phase." They demonstrate technical feasibility, refine process understanding, and build internal expertise. However, expecting these pilots to scale seamlessly into complex end-to-end processes within the next one to two years is unrealistic. Second, an honest assessment of platform strategy is crucial. Companies must determine which orchestration layer best integrates with their existing IT infrastructure. Those centered on a specific business application platform will need to monitor its orchestration offerings closely. Hyperscaler-centric architectures will lead to different conclusions. Hybrid scenarios, common in mature B2B environments, will present the most complex challenges, and this strategic discussion cannot be delegated. Third, documenting processes in an "agent-ready" format is an immediate and critical preparation. Regardless of which orchestration layer ultimately prevails, well-defined processes with clear handovers, documented decision logic, and established escalation paths are prerequisites for effective automation. Neglecting this step will necessitate rebuilding process definitions once more advanced tools become available during the "Windows phase." ### What remains Transition phases are typically easier to identify in retrospect than in real-time. The challenge lies in recognizing these shifts prospectively without succumbing to hype. Those who dismissed DOS in the early 1990s as "unusable" missed a significant technological leap. Similarly, dismissing today's agents as "not yet ready" could lead to a comparable oversight. Conversely, declaring every agent pilot a production-ready business process mistakes the current "DOS phase" for the future "Windows phase." The fundamental principle remains: prioritize processes over tools, especially during this transitional period.