Agentic AI for Science: From Chatbots to Autonomous Research Assistants

Science is where agentic AI gets interesting. A chatbot can explain what density functional theory is; an agent can set one up, run it, notice the calculation did not converge, fix the input and report the energy. Over the last few years my colleagues and I have worked on software for scientific automation: self-driving laboratories first, and more recently AI agents for quantum simulation and quantum chemistry. This article explains why science suits agents so well, what is hard about it, and what we have learned.

Why science suits agents

• The tools already exist. Simulators, databases, analysis scripts and lab instruments are all programs with inputs and outputs, exactly what an agent can call. (If tool use is new to you, see how agents use tools.)
• Feedback is often objective. A calculation converges or it does not. An energy is consistent with a reference or it is not. That makes self-checking possible.
• The workflow is long and fiddly. Expert time goes into software syntax, file formats, parameter choices and babysitting jobs, rather than into scientific thinking. Automating that is valuable.
• The accessibility gap is large. Powerful methods remain in the hands of specialists because the software is complex. An agent that translates intent into correct calculations can widen the circle of people who use them.

What is hard

• Structured state. Geometries, wavefunctions, tensors and trajectories are large and numerical. They do not belong inside a language model’s context window.
• Reproducibility and provenance. A scientific result needs a record of exactly what was run, with which inputs and which versions.
• Validation. A fluent but wrong answer is worse than no answer. Agents must check their work against physical and numerical constraints.
• Heterogeneous software. Every package has its own conventions, and documentation is the real specification.
• Safety and cost. Calculations burn compute, and laboratory actions have real-world consequences.

Four systems from our group

The “El Agente” family, from the Aspuru-Guzik group at the University of Toronto, tackles these problems in different domains. I am a co-author of the first three.

The descriptions above paraphrase the papers’ abstracts; see the papers for details and results. Together they illustrate a design question that matters well beyond chemistry: should the work be split across many cooperating agents, or run by one agent over structured state? We discuss that trade-off in multi-agent systems.

From software agents to self-driving laboratories

The same ideas extend from simulation to the physical lab. A self-driving laboratory closes the loop: software plans an experiment, robots carry it out, instruments measure the result, and the outcome informs the next plan. Two of my papers touch this:

• ChemOS 2.0 (Matter, 2024) is an orchestration architecture that treats the laboratory as an “operating system”, coordinating communication, data exchange and instructions among modular lab components, and combining ab initio calculations, experiment orchestration and statistical algorithms to guide closed-loop operation.
• The digital pipette (Digital Discovery, 2023) is a low-cost, 3D-printed open-hardware pipette designed so that robots can do liquid handling, a basic operation of any chemistry lab.

An agent that can plan and run calculations, plus an orchestration layer that can run experiments, points towards research assistants that handle more of the loop between hypothesis and result.

Design principles that keep coming back

1. Keep numbers out of the prompt. Pass structured scientific state by reference, as typed objects or graph nodes, not as pasted text.
2. Record provenance by default. Every result should be traceable to its inputs, code version and decisions.
3. Let documentation be the interface. Agents that read library docs generalise to new software far better than agents with hard-coded recipes.
4. Validate against physics. Check convergence, units, symmetries and reference values, and make the agent report failures instead of hiding them.
5. Keep the scientist in charge. The goal is a collaborator, not an oracle; humans approve expensive or irreversible steps and interpret the results.
6. Evaluate honestly. Build benchmarks from real tasks and measure reliability across repeated runs. See how to evaluate and secure AI agents.

Where this is heading

Agentic systems will not replace scientific judgment. They are likely to change who can run sophisticated computations, how quickly a hypothesis can be tested, and how well computational work is documented. I find that exciting, and I will keep writing about what works and what does not. For the full list of my research, see the research page and the publications.

New to agents? Start with What is Agentic AI?