Agricultural innovation systems (AIS) are increasingly recognized as complex adaptive systems in which interventions cannot be expected to create predictable, linear impacts. Nevertheless, the logic models and theory of change (ToC) used by standard-setting international agricultural research agencies and donors assume that agricultural research will create impact through a predictable linear adoption pathway which largely ignores the complexity dynamics of AIS, and which misses important alternate pathways through which agricultural research can improve system performance and generate sustainable development impact. Despite a growing body of literature calling for more dynamic, flexible and "complexity-aware" approaches to monitoring and evaluation, few concrete examples exist of ToC that takes complexity dynamics within AIS into account, or provide guidance on how such theories could be developed. This paper addresses this gap by presenting an example of how an empirically-grounded, complexity-aware ToC can be developed and what such a model might look like in the context of a particular type of program intervention. Two detailed case studies are presented from an agricultural research program which was explicitly seeking to work in a "complexity-aware" way within aquatic agricultural systems in Zambia and the Philippines.