When AI Takes Over Research, What Role Remains for Human Scientists? Inside AgentSociety²

AgentSociety², developed by a Tsinghua University team, is positioned as an "Integrated Research Environment for Executable Social Science" that embeds AI Social Scientists within a human‑in‑the‑loop workflow. The system expands the scope of searchable problems, reduces engineering overhead, organizes experiments, and compiles results, while humans keep control at critical decision points.

Three‑layer design underpins the platform:

Layer 1 – Dual‑role closed loop : AI agents (silicon participants) and AI Social Scientists share the same execution environment. Silicon participants act as "silicon participants" within large‑scale social simulations, while AI Social Scientists orchestrate literature review, hypothesis formation, experiment design, simulation execution, analysis, and drafting.

Layer 2 – Human steering : The platform adopts a human‑in‑the‑loop approach. AI reduces the engineering burden by configuring experiments, running simulations, and generating reports, but humans retain judgment over hypothesis validity, variable mapping, intervention design, and interpretation of results.

Layer 3 – Mechanism laboratory : Research questions and theoretical hypotheses are translated into executable configurations—agent behaviors, environment rules, interventions, and metrics—so that social‑science mechanisms become runnable experiments rather than static textual statements.

The harness layer connects skill libraries, sub‑agents, tool interfaces, and staged workflows, ensuring that AI actions are bounded by explicit research stages and can be inspected or corrected by researchers at key checkpoints.

AgentSociety² employs a skill‑based agent architecture: observation, cognition, planning, memory, and domain‑specific decision rules are modularized, and each agent maintains an independent workspace with state, memory, logs, and checkpoints, enabling long‑term, traceable behavior in massive simulations (supporting tens of thousands of agents).

Four sequential AI‑driven phases structure the research pipeline:

Literature‑driven hypothesis generation : AI extracts theoretical background, research gaps, and variable relationships from literature, packaging them into structured hypothesis bundles with cited evidence.

Semantic‑to‑executable configuration : AI translates research questions into simulation‑ready configurations, preserving high‑level semantics while generating concrete initialization files, experiment steps, intervention conditions, and evaluation metrics.

From simulation results to mechanism analysis : After execution, AI aggregates numerical outcomes, interaction logs, and context data, mapping them back to the original hypotheses to compare group behaviors and emergent patterns.

From analysis report to draft paper : AI assembles the research theme, hypotheses, experiment setup, analysis summary, figures, and references into a draft manuscript, aligning claims with evidence and performing multi‑agent checks for consistency before handing the draft to the human author.

AgentSociety² demonstrates its generality through seven multi‑scale social‑science experiments covering micro‑level behavioral studies, meso‑level network dynamics, and macro‑level urban and disaster scenarios. These include public‑goods games, prisoner's dilemma, opinion polarization, information bubbles, and city mobility simulations, illustrating how the platform can accelerate research cycles from weeks to hours.

By integrating literature review, hypothesis management, simulation, analysis, and writing into a single auditable system, AgentSociety² shifts social‑science research from labor‑intensive pipelines to a rapid, reproducible workflow, while preserving the essential scientific judgment that only human researchers can provide.