Agents for Experiment, Experiments for Agents: A Topological Framework for Automated Mechanism Discovery

The transition of AI from static predictors to strategic agents has elevated mechanism design—the engineering of interaction rules, authority flows, and feedback loops—to a central challenge in algorithmic governance. While experimentation serves as the primary instrument for uncovering these dynamics, the traditional manual design process cannot keep pace with the combinatorial complexity of agentic systems. We propose a recursive paradigm of Agents for Experiment, employing AI architects to design the protocols used to study AI systems (Experiments for Agents). However, without a rigorous ontology, current agentic approaches rely on unstructured text, potentially yielding operationally invalid or scientifically trivial designs. To bridge this gap, we introduce SEED (Structural Encoding for Experimental Discovery), a framework that formalizes experimental protocols as computable runtime execution graphs. By decoupling the topological skeleton of an interaction from its semantic context, SEED provides a unified grammar for automated discovery. We demonstrate the framework's utility through three distinct functions: (1) Descriptive Utility, which synthesizes fragmented literature into a standardized topology library; (2) Evaluative Utility, which operationalizes scientific novelty via computable evaluation scores; and (3) Generative Utility, which enables a "Generative Topology Search" algorithm. This allows agentic researchers to systematically identify structural gaps, which are counterintuitive governance architectures adjacent to established science, and propose novel experimental designs. We conclude that SEED transforms mechanism discovery from an artisanal craft into a structured optimization problem, laying the foundation for high-throughput experimental science in the agentic era.