A new academic paper offers a map for one of AI's messiest engineering problems: how to actually build an agent.
Researchers behind the paper introduce a taxonomy for foundation-model-based agents, cataloging the architectural options available for both what an agent can do, its functional capabilities, and how well it does it, meaning non-functional qualities like reliability or latency. The paper breaks the agent lifecycle into design-time and run-time phases and lays out the tradeoffs at each stage. On top of the taxonomy, the authors propose a decision model meant to guide developers through specific architecture and runtime choices rather than leaving them to trial and error. The paper frames itself explicitly as a response to what its authors call fragmentation in the field.
That fragmentation is real. Agent-building right now is a patchwork of frameworks, each with its own vocabulary for memory, tool use, and planning, which makes it hard to compare designs or reuse lessons across projects. A shared taxonomy, if it catches on, could give teams a common language for architecture reviews and a checklist for decisions they are currently making ad hoc.
Whether it catches on is the open question. Taxonomies are cheap to publish and expensive to get adopted, and the field is littered with proposed standards that never made it past the paper that introduced them. This one's value will depend less on its logic, which looks sound, and more on whether framework builders actually start speaking its language.