A Taxonomy of Architecture Options for Foundation Model-based Agents: Analysis and Decision Model

In the last few years, a new kind of artificial intelligence has emerged, one that does not just answer questions but takes action. These are not simple programs that follow a rigid list of instructions. Instead, they are systems built on massive "foundation models," which are like vast libraries of human knowledge and language that have been trained to understand context and generate new ideas. When developers connect these models to tools that can browse the web, write code, or control software, they create "agents." These agents can perceive their environment, reason through a problem, plan a sequence of steps, and act to achieve a goal, much like a human employee might. From managing complex software projects to navigating virtual worlds, these systems promise to revolutionize how we work. Yet, as these agents become more capable, the way engineers build them has remained chaotic. Without a shared language or a clear map of choices, designers often piece together systems in an ad-hoc fashion, making decisions based on guesswork rather than a systematic understanding of how different parts fit together and what trade-offs they create.

A team of researchers set out to bring order to this growing complexity. They conducted a massive review of 66 scientific studies to understand exactly how these foundation-model-based agents are constructed. Their goal was not just to list the parts, but to create a structured guide that explains the consequences of every design choice. The result is a comprehensive "taxonomy," which is essentially a detailed classification system that organizes the many ways an agent can be built. They found that architects must make fundamental choices about how an agent receives information, how it remembers past events, how it plans its actions, and how it coordinates with other agents. For instance, an agent might be designed to work alone or as part of a team, to remember everything it sees forever or to forget old details to save space, and to follow a single straight path to a goal or to explore multiple possible routes simultaneously.

The researchers discovered that every one of these choices carries a cost. There is no perfect design that excels at everything. If an architect chooses a system that is highly flexible and can adapt to unexpected changes, they often sacrifice the ability to easily test and fix that system later. If they prioritize a memory system that is fast for the computer to process, they might make it harder for a human to understand why the agent made a specific decision. To help engineers navigate these difficult compromises, the team built a "decision model." This model acts like a map, showing exactly how picking one architectural option impacts the system's reliability, speed, security, and ease of use. It makes the hidden trade-offs visible, allowing designers to choose a path that aligns with their specific needs, whether that is building a system that must never fail or one that can learn and adapt on the fly.

To ensure their findings were useful and accurate, the researchers tested their framework with twelve experts from both universities and major technology companies. These experts, who have years of experience building complex AI systems, reviewed the classification and the decision model. The experts agreed that the framework successfully captured the essential building blocks of these agents and provided a clear, practical way to reason about their design. They found that the model helped clarify choices that were previously confusing, such as how to balance the need for an agent to be autonomous with the need for it to be safe and predictable. The study confirms that the field is moving away from random experimentation toward a more disciplined engineering approach, where decisions are made with a clear understanding of the consequences.

The work does not claim to have solved every problem or to provide a single best way to build an agent. Instead, it offers a structured foundation for the next generation of design. By organizing the landscape of options and explicitly mapping out the trade-offs, the researchers have given architects a common vocabulary and a rigorous method for decision-making. This shift allows the field to mature, moving from a collection of isolated experiments to a cohesive discipline where systems can be designed with intention, reliability, and a clear understanding of their limitations. As these intelligent systems become more integrated into our daily lives, having a clear map of how they are built and what they can and cannot do will be essential for creating technology that is both powerful and trustworthy.