To achieve the benefits of human-AI teams, we need to ensure sophisticated infrastructure, human oversight, strong governance and a culture agile enough to shift from command to orchestration, argues Jan Bosch.
As we’re taking ourselves through the R&D maturity model, we’ve reached the AI supercharger stage. This is where most of the team is substituted by AI agents. The idea is that an individual, or maybe a pair of people, supported by a set of AI agents, replaces the work of one or more Agile teams. Here, the real productivity gains of AI start to materialize.
As an example, in one of the cases shared during the interviews, an enterprise architect was teamed up with a set of AI agents that dealt with generating requirements, designing the architecture, producing the code and test cases, as well as conducting integration and testing. The role of the architect was predominantly to provide the right prompts to the agents and to supervise their output.
Of course, this change isn’t realized just by saying it out loud. Instead, companies need to get several aspects in place to capitalize on it. First, they need to identify the roles and capabilities of the one or two humans working with the AI agents, as these people need to be generalists who can supervise all aspects of an R&D project. Second, the interface between the humans and the agents needs to be clear so that the result is of sufficient quality. Third, this is also the stage where ways of working need to be redesigned. Fourth, the prerequisites, in terms of infrastructure and organizational readiness, need to be in place. And, finally, the risks associated with this approach need to be understood and managed.
At the heart of the AI supercharger model is a small team, often just one person, or perhaps two, paired with an ensemble of AI agents. These human orchestrators must be generalists, capable of overseeing every phase: from ideation and architectural design to coding, testing, integration, deployment and beyond.
These individuals shift from hands-on execution to orchestration, including prompting, evaluating and validating AI outputs. Their value lies not in writing every line of code, but in defining intent, ensuring alignment and managing complexity at a system level. As is sometimes said, AI agents are like genies: they may grant your wish, but often in unpredictable ways. Managing these agents safely and effectively requires high-level thinking, vision and general competence in all areas of product development.
Although the individual has the potential to be an order of magnitude more productive, it requires a robust and transparent interface between human and agent. We need frameworks for managing prompts as well as a shared project memory and clear feedback loops so agents remain context-aware. For example, retrieval-augmented generation (RAG) can be used to ensure that agents maintain context. Also, agents must be trained – or “onboarded” – to ensure that the governance guidelines, ethical guardrails and security boundaries are complied with in the same way humans need to follow them. In many ways, engineers become more of a manager supervising AI colleagues rather than executing tasks themselves.
As with the maturity model focusing on the business processes, we also need a “zero-based thinking” redesign of ways of working for R&D. The AI supercharger workflows are different from traditional Agile practices. For instance, planning is no longer based on sprint-based backlogs but rather focuses on prompt-oriented tasks, where describing intent in plain language replaces user story creation. Also, Agile retrospectives and coordination rituals are replaced with continuous, AI-driven feedback loops. Agents may generate tests, debug, deploy and even retrospect automatically with engineers providing oversight. Although it may seem easier to “just provide oversight,” in the end, we build software and products for other human users, which requires us to ensure that the quality, value and user experience are satisfactory. This is difficult for agents to realize.
One of the factors often underestimated when exploring the supercharger stage is the need for enablers and infrastructure. We need a reliable and modular AI infrastructure that supports multi-agent coordination, retrieval-augmented workflows and safe execution. Also, even though we may cease to use traditional Agile ways of working, we still need a structured development lifecycle, but one that’s focused on AI, including design, training, validation, deployment and continuous monitoring of agents. As the AI models are continuously improving, both our infrastructure and our lifecycle need to be modular so that it’s easy to replace one agent with another. Finally, we need to be careful in managing risks such as hallucinations, security threats and the erosion of institutional knowledge. As we increasingly rely on highly autonomous agents, these risks become increasingly prevalent and need to be managed.
As the saying goes, if something sounds too good to be true, it typically is. Here, it’s no different. We’ve mentioned risks a few times already, but this needs to be a significant area of attention. Some research shows that agent-generated code may result in a significant increase in duplicate code, which may raise maintenance cost and technical debt. Second, agent-generated code may decrease delivery stability as some studies seem to suggest. Third, although many are jumping on the AI bandwagon, there are indications that a significant percentage of AI agents may be decommissioned for failing to deliver significant benefits.
For all the talk about risks, though, there are several claims out there that show that if you get it right, there can be quite significant benefits. For example, Salesforce replaced 4,000 support roles using AI agents responsible for breaking down and executing complex tasks. Some publications claim that AI agents are raising developer productivity by up to 30 percent, letting engineers focus more on strategic design and less on boilerplate coding. Github Copilot experiments show developers completing tasks nearly 56 percent faster and trials within Google, especially for routine work, yield around a 21 percent reduction in time on task. Some research shows that human-AI teams demonstrated 60 percent greater productivity per person for creative tasks.
The AI supercharger stage represents a major shift: one human or a duo empowered by AI agents can match or be more productive than one or more Agile teams. To achieve these outcomes, though, we need to ensure sophisticated infrastructure, human oversight, strong governance and a culture agile enough to shift from command to orchestration. Rather than treating AI agents as tools, they increasingly become teammates, allowing us to unlock and supercharge human capability. To end with a quote by Satya Nadella: “AI agents will become the primary way we interact with computers in the future. They will be able to understand our needs and preferences, and proactively help us with tasks and decision-making.”
