Socratic AI Agents Achieve Autonomous Scientific Discovery in Complex Systems

The field of automated scientific discovery is advancing, with AI systems now capable of operating instruments and generating hypotheses. However, true autonomy requires "epistemic autonomy"—the ability for AI to construct, challenge, and revise its own scientific explanations. This research presents AHOIS, a multi-agent AI scientist designed to embed Socratic midwifery into experimental processes. AHOIS features a "physics-critic" agent that rigorously interrogates hypotheses. This agent employs causal questioning, constraint checking, counterexample generation, and the formulation of falsification criteria to ensure robust scientific reasoning. The system was evaluated on a complex, high-dimensional multimode-fibre optical platform, which presents challenges like wave transformations, indirect detection, and environmental drift. Remarkably, AHOIS autonomously proposed and validated a random-interference encoding hypothesis, discovered adaptive sparse-measurement strategies, and diagnosed various failure modes without any pre-encoded schemes or models. It even translated a published imaging protocol into an executable workflow on a different setup. This demonstrates a significant step towards self-correcting, evidence-grounded autonomous discovery in intricate physical environments.