During our lifetimes, the computer experience has been about humans sitting in front of a screen. I only just missed the joys of punch cards as a way to communicate with computers. By the time I started working, in 1983 as a systems programmer for Esso, the last punch card reader at the company had been retired just a year before, and I began my career sitting in front of a computer terminal typing instructions on a keyboard. Still, the way to tell a computer to do things was very much based on typing commands into a command line. IBM mainframes used “job control language” (JCL), the first computer language I learnt, and this was essentially a series of 80-column lines in a file, designed to mimic punch cards. The IBM PC was launched in 1981, and this still had a command-line-interface; it took several years before a graphical interface with a mouse appeared. It was only in 1984 that Apple launched the Macintosh computer, with its graphical user interface and mouse. That itself was inspired by research at Xerox PARC, which Steve Jobs had visited in 1979. Microsoft Windows 1, with its tiled windows and drop-down menus, appeared in 1985. The next major evolution was touchscreen phones (the iPhone in 2007) and tablets like the iPad (2010). More recently there have been further advances, such as speech recognition and virtual reality.
Software vendors have spent considerable effort refining the user interface, but as of today the major enterprise applications (ERP, CRM, supply chain, etc.) deployed around the world still rely mostly on a series of graphical menus. The advent of AI allowed for some new possibilities. Natural language interfaces and “copilots” give users new options but don’t fundamentally change the nature of the user interface. Instead of navigating a series of screens, a user might type “show me sales in Germany last quarter” into a chat interface and get an answer. This may be an improvement, or at least be more flexible, but is not in itself a complete change in how we work.
It is different with AI agents. An AI agent is not interested in where the “book” button on the main screen of a reservation system is. An agent would prefer to interact with systems via an application programming interface (API), which is a communication interface between computers. An agent will use an API to securely authenticate with an application or service and will then use APIs to, for example, gather prices of flights from multiple airline websites, or the prices of hotels from multiple hotel websites for comparison.
In a world of agents, the classic human user interface is less relevant. Instead, the quality of the API matters. How stable is it? Does it have clear metadata and permissions? Does it have good documentation and predictable error handling?
Traditionally, each agent-to-application interface required some level of bespoke work, which is time-consuming. The emerging open-source standard Model Context Protocol (MCP) gives AI applications a universal way to connect to external data sources and tools. A program translates data from a specific tool, like a SQL database or Google Drive, into a standard interface that an AI application can use. Think of it like a USB port for AI, allowing different AI applications to connect to tools without needing a custom integration for each one. There are a number of issues with MCP, such as authentication, performance, and indirect prompt injection, but that is a separate topic. It has been rapidly adopted.
This gradual move to a world of agents and APIs has implications for application software vendors. Instead of dealing with a world of human users, vendors have to handle a world where AI agents may never see a user interface, and instead interact via an API. If users end up just invoking a series of agents, this effectively disintermediates the software vendors: users may no longer see the application’s user interface and brand. This means software developers now have to handle two audiences: humans and AI agents. A product with a highly effective, well-documented and efficient API may out-compete a rival that has an attractive GUI but a poor API. It should also be noted that cutting out the user interface does not remove the need for good data management. Ambiguous metadata, inconsistent master data terminology or poor-quality data are every bit as much of a problem for an AI agent as they are for a human user. Humans will still need their dashboards and visualisations, if only to check what their agents are up to, but the application software vendors now have two parallel audiences to cater for: human and agent.
After decades of making computers easier for humans to understand, the next major software evolution may be about making software easier for computers to understand.
