[2609.31219] Research with AI Agents: How Agentic Systems Are Changing Scientific Work

Agentic AI systems are shifting scientific work from manual execution to strategic steering and review, allowing researchers to focus on high-value decision-making and ethical oversight. These systems independently decompose complex tasks—such as literature searches, data analysis, and programming—into subtasks, enabling high-speed digital research while reducing human workload.

Key transformations include:

• Task Design and Execution: Agents automate repetitive workflows and experimental protocols, with efficiency gains being greatest when behaviors can be formalized and tested automatically.
• Evaluation and Review: While agents handle initial analysis, human researchers remain responsible for validating results, interpreting significance, and ensuring ethical standards, as automated explanations often lack reliability.
• Knowledge Curation: Agents streamline literature reviews and data synthesis, freeing up time for experts to assess the societal and clinical relevance of findings, which cannot be fully determined by formalized criteria.

Despite these gains, challenges remain in reviewability; complex agentic systems produce extensive reasoning traces that are difficult for humans to audit, necessitating robust component validation and human-in-the-loop oversight to maintain scientific rigor.

Research with AI Agents: How Agentic Systems Are Changing Scientific Work examines agentic AI systems as a new layer in scientific production, moving beyond single-shot LLM assistance toward systems that can plan, call tools, manipulate data or code, iterate on results, and maintain task memory. The material organizes this transformation around four core workflow dimensions: task design, execution, evaluation, and knowledge curation. In task design, agents can help translate ambiguous research questions into decomposed, testable subproblems, propose hypotheses, or generate experimental and computational protocols. In execution, they can coordinate code, simulations, databases, and potentially laboratory instruments, enabling longer-horizon research workflows that were previously too labor-intensive to automate end-to-end.

A central contribution of the work is its focus on the governance and epistemic challenges introduced by scientific agency. It distinguishes between narrow copilots and more autonomous research agents, and argues that reliability depends not only on model capability but on orchestration, verification, provenance, and human oversight. Because agents increasingly produce the artifacts scientists review—code, analyses, literature syntheses, and experimental plans—evaluation must move beyond isolated task accuracy to end-to-end workflow quality. The material highlights failure modes such as hallucinated citations, overconfident interpretation, unsafe actions, and reproducibility gaps, while also emphasizing the importance of audit trails and domain-specific benchmarks.

The work matters because it anticipates a shift in research infrastructure and scientific norms. If agentic systems become standard in discovery, science will need new interfaces, evaluation methods, and accountability mechanisms that make agent behavior inspectable and scientifically trustworthy. It frames the key question not as whether agents can perform individual tasks, but how research institutions should integrate them while preserving rigor, reproducibility, and responsible knowledge curation.