Paper #11: Adrasteia: AHuman-in-the-Loop Agentic Workflow for Theory-Oriented Research in Economics

We introduce ADRASTEIA, a human-in-the-loop agentic workflow for theory-oriented research in economics, with only cautious extension to ad jacent theory-oriented settings. Unlike mathe matical reasoning systems, which operate pri marily through proof search under strong veri fication, theoretical research in economics often proceeds through assumption search: identify ing assumptions and model structures that are coherent, meaningful, and disciplinarily accept able. ADRASTEIA is designed for this setting. It organizes research into a modular workflow consisting of problem framing, assumption pro posal, internal verification, revision with alter native branching, human adjudication, and ex ternal validation through simulated review. A central design principle is that human interven tion is not a fallback for model failure, but a mechanism for preserving originality, interpretive control, and scholarly accountability. We argue that this workflow provides a plausible founda tion for AI-assisted theory-oriented economics beyond one-shot prompting or proof-oriented rea soning paradigms, and we report pilot evaluation evidence organized around a primary automated proxy comparison, secondary diagnostic compar isons, and exploratory blinded pairwise human evaluation. The resulting empirical picture is de liberately limited: the current evidence is con sistent with moderate gains in a reviewer-style proxy layer, but it does not constitute independent validation of research-system superiority