Lawrey E1
1Australian Institute of Marine Science, , Australia
Biography:
Dr Eric Lawrey joined the Australian Institute of Marine Science (AIMS) 18 years ago to develop the eAtlas, a repository for research data. He now leads the Knowledge Systems team, developing information systems that help researchers and reef managers access environmental research data. Eric has extensive experience with spatial data analysis and remote sensing, mapping the reefs of the Torres Strait, the Keppel Islands and the Coral Sea. Before AIMS he led a team developing new methods for software engineering. He studied Computer Systems Engineering and completed a PhD in wireless communications, modelling techniques now used in Wi-Fi and digital television.
Abstract:
Citing the literature is essential to research writing but slow. A reference manager records what an author has read, yet over months or years the catalogue grows too large to recall. Writing an introduction then means searching the catalogue by hand, which takes considerable time.
We built an agent workflow to speed this up. It takes a passage of text, identifies the statements that need a citation, searches a Zotero reference library for likely matches, and recommends sources. The author confirms which statements to cite and which recommendations to accept. The goal is to reduce the checks the author must make.
An agent workflow can be built in many ways. We built and tested the same workflow in four forms, from least to most structured. The least structured asks a frontier model in a generic agent harness to use the reference manager directly. A skill file then supplies tested strategies and code fragments. A further step adds Python scripts the agent calls as tools, handling the deterministic work: a cached copy of the library and semantic search from an embedding model. The most structured embeds the whole workflow in a desktop application that calls a language model only where judgement is needed.
Preliminary testing shows a clear trade-off. Added structure allows smaller models to complete the workflow, but at the cost of reduced adaptability. From this case study we share what we learned about matching agent structure to model capability, development effort, and the accuracy a task needs.
