Agentic AI
Agentic AI refers to systems, usually built around large language models, that act with some independence: they interpret observations and instructions, plan multi-step actions, call external tools such as software, databases or laboratory hardware, and revise their plans from the results. In science this ranges from literature agents and retrieval-augmented question answering to frameworks that run a research cycle from hypothesis to manuscript, and to agents that operate instruments in self-driving laboratories. A useful distinction is whether the model sits inside a closed reasoning or control loop (agentic) or is only called as a component, for example to predict a recipe, within an otherwise conventional automated workflow.
Agentic science as a paradigm (2025)
Xin, Kitchin and Kulik argue in a Comment that agentic science, in which AI agents explore and learn with partial or full autonomy, may be the next paradigm after empirical, theoretical, simulation-based and data-driven science (Xin et al. 2025). They survey systems such as the AI Scientist, FutureHouse, LLaMat, Coscientist and LLaMP, point to the Model Context Protocol and Agent2Agent protocol as steps toward interoperable multi-agent systems, and expect the earliest benefits in data-rich, automated fields such as chemistry and materials science. The challenges they emphasize are hallucination and its compounding over long tasks, the lack of standard evaluation metrics (pass@k, passk, step counts and reasoning-graph distances have been proposed), and prompt fragility. Their remedies are grounding in knowledge graphs, verification agents or human review at decision points with automatic halting of implausible plans, reporting of model versions, prompts and agent dialogues for reproducibility, and checks against bias such as training data that include negative results. Summary: xin2025.
Agents for laboratory instruments (2025)
Kitchin showed in practice the building blocks of such agents (tool calling, code generation, and a program loop with memory) applied to the Claude-Light instrument, along with the risks of generated code and nondeterministic output (Kitchin 2025). Summary: kitchin2025b.