Type: book
Abstract: An agent is a system in which a model chooses the control flow at run time. Everything difficult about building one follows from that single property, because the set of reachable states stops being enumerable and test coverage stops meaning anything. This book catalogues thirty patterns for building agents that keep working, organised around failure evidence. Five recurring failure modes are consolidated from published incident analyses, among them a taxonomy derived from more than 1,600 annotated execution traces across seven frameworks. Every pattern states the failure it addresses and the price it charges in tokens, latency, or capability. The literature surveyed here reaches an uncomfortable conclusion. Better base models will not remove these failures, since they are properties of systems that let a model decide what happens next. They have to be engineered against. Every quantitative claim is attributed to a named source, and all 95 references were verified against primary sources.
Author: Gaikwad, Madhava
Language: English
Subfield: Software
Field: Computer Science
Domain: Physical Sciences
Open Access status: green
OpenAlex ID: W7213382365