When OpenAI introduced Sites and role-specific plugins for Codex, much of the discussion focused on the user experience. Users can now instruct an agent to build interactive workspaces through an agent conversation, connect them to enterprise systems, and tailor functionality to specific business roles. While the announcement was framed as a product update, it points toward a larger shift that has been quietly unfolding across enterprise software over the past two years.
The traditional relationship between users and software is beginning to change.
For decades, enterprise applications existed as relatively fixed artifacts. A business identified a need, requirements were collected, engineers designed a solution, and the resulting application was deployed for employees to use. Even low-complexity internal tools followed some version of this process. The software was created first, and only afterward did users interact with it.
Agent-based systems blur that boundary. Instead of navigating a predefined application, users describe an outcome, provide context, and allow software to be assembled dynamically around the task itself. The distinction may seem subtle, but it represents a significant departure from how enterprise systems have traditionally been built.
From applications to generated workspaces
Most organizations maintain dozens, if not hundreds, of lightweight operational tools that were never intended to become products. Sales teams need reporting dashboards. Operations teams require launch management workspaces. Customer success departments rely on onboarding trackers and escalation portals. Finance teams continuously request interfaces that combine information from multiple systems into a format suitable for decision-making.
The technical complexity of these tools is often modest. Their cost comes from the coordination required to build and maintain them. Product managers gather requirements, engineers allocate time, infrastructure teams provision resources, and security teams review access controls. A dashboard that might ultimately consist of a few database queries and visualizations can spend weeks moving through organizational processes before anyone uses it.
What platforms like Codex suggest is that a growing category of enterprise software may eventually bypass much of this cycle.
Why orchestration is becoming more important than models
Over the past several years, AI discussions have largely centered on models. Organizations compared GPT-4 against Claude, evaluated context windows, measured benchmark performance, and debated which provider offered the strongest reasoning capabilities. While these comparisons remain important, they overlook the part of the system that determines whether enterprise AI delivers meaningful value.
An AI system rarely struggles because it lacks access to the right information, cannot retrieve relevant context efficiently, does not understand organizational permissions, or has no mechanism for interacting with business systems. As a result, many of the most important engineering challenges in enterprise AI have shifted away from model selection and toward retrieval, orchestration, governance, and integration.
Role-specific plugins directly address this problem. They transform enterprise systems from passive sources of information into active participants within agent workflows. Instead of copying information between applications, employees can rely on an agent conversation that operates across systems while respecting existing permissions and business logic.
Viewed from this perspective, Sites become far more interesting than a low-code website builder. What OpenAI is effectively introducing is an abstraction layer capable of transforming structured intent into usable software. The more important capability lies in connecting user requests to enterprise data sources, workflows, and actions while producing a coherent environment through which those interactions can occur.
Infrastructure evolved from physical servers to cloud platforms. Deployment pipelines replaced manual releases. Databases became managed services. Authentication became an API. At each stage, engineering effort moved further away from implementation details and closer to architectural concerns.
Rather than spending time constructing interfaces, engineering teams may focus on defining system boundaries, access models, governance frameworks, and data contracts. The responsibility shifts from building every application individually toward creating environments in which applications can be generated safely and predictably.
Organizations that successfully adopt these systems will require robust identity management, clear permission structures, comprehensive observability, and well-defined data governance policies. Without these foundations, AI-generated applications risk becoming a new source of operational complexity rather than a solution to it.
The broader implication is that enterprise software may be entering a period in which application creation becomes fluid. For a substantial category of internal tools, the path from identifying a business problem to deploying a working solution may no longer involve a traditional development lifecycle. Instead, software will emerge from the interaction between users, an agent conversation, organizational context, and enterprise systems.
Whether Codex ultimately becomes the platform that enables this shift is less important than the direction it reveals. The most consequential change is that software itself is beginning to behave less like a static product and more like a dynamically generated resource, assembled at the moment it is needed and shaped around the task at hand.
