As we move from using AI agents to systems of fully autonomous and generators, humans are more interested in asking the right questions than building solutions, says Jan Bosch.
After the first three steps of the business process maturation ladder and the R&D maturity ladder, i.e. AI assistant, AI compensator and AI supercharger, we will discuss the fourth level, the AI system generator. Here, the intention is to fundamentally change from extending the role of humans to creating a fully autonomous end-to-end system. It really doesn't matter how things work behind the scenes (when was the last time you checked out the inside of the compiler you use to build your code?), but it's important to note that in most cases, the AI system generator is not a single agent, but a set of tuning agents that replicate the complete lifecycle of the software engineering process.
Of course, there are people involved in the development of the system's intent, but the intent is an input to the AI system generator, and all of the points are done autonomously by the agent. Once a system is generated, humans can provide input in the form of additional, complementary, or modified intent. This will cause the generator to play all or part of the system. Some people think it's virtually impossible for AI agents to create such a system, but anyone who has seen systems like Lovable and its many competitors knows they're heading towards a “no code” solution where users do this accurately.
While it may not be strictly necessary for noncritical systems, most people want to review prompts or intents, high-level design of the system, test cases for test components, and requirements generated based on the entire system and documentation and deployment process. The proof lies in the pudding, or deployed running system, but often some guarantees are required in line with the way the generated system performs to actually run.
In other words, the workflow that the AI system generator follows follows follows many of the steps that appear in the human development process. First, intentional capture requires natural language processing to ensure that the underlying organizational needs are actually properly recorded.
Second, systems need to be structurally decomposed from an architectural and component perspective, and from the perspective of the tasks they need to perform. In this regard, it is important to note that components must be taken and configured from open or commercial sources. Alternatively, you need to generate code for the component. Finally, of course, it is perfectly feasible to use machine learning or deep learning models to achieve functionality.
Third, AI system generators need to generate tests of all types: units, integrations, systems, to ensure that the generated system works as intended. You will also need to generate the reports, user manuals and artifacts needed to comply with and certify the regulations.
Finally, the generated system must be delivered. It is preferably deployed directly, similar to today's DevOps pipeline, but in many cases companies want people in the loop for the final step towards deployment. However, the more automated the process, the smoother the workflow will be, especially when the system is played regularly and partially or completely completely regenerated.
The advantages of this approach are clear. The speed at which a system can be generated is improved by an order of magnitude. It is not uncommon for small systems to be generated in minutes or hours. Even if it takes a few days, it's infinitely faster than the traditional approach. Furthermore, there are much less human error, ideally, the way you work with the system is very repeatable, with much more consistency and quality. And of course, it provides enormous strategic leverage as we humans can focus on what we humans build, why we build it, and how to do it to AI system generators.
For all the pluses, we need to be aware of some challenges. The automatically generated system should be able to actually depend on the company as much as it needs. Additionally, in the context of regulatory compliance, e.g. safety, security, or other aspects, you should ensure that the generated artifacts cover these needs. This makes the system manager accountable if it is generated by AI. Ultimately, despite the role that changes from builder to supervisor, it must be a company person.
Moving from using AI agents to an extension to a fully autonomous generation system, humans become supervisors, verifiers, and strategists. They are more interested in asking the right questions than building a system. In a world where AI generates systems, you need to “take an hour to solve a problem, think about 55 minutes on a problem and 5 minutes on a solution, and spend a fifth hour thinking about a fifth hour and focus your energy on the intention of framing rather than the solution.”
