How AI Does Science — SciencePal

Ask a working scientist where their week went, and a surprising amount of the answer has nothing to do with science. It is plumbing: installing software, converting file formats, gluing one tool's output into another tool's input, debugging a broken environment, re-running a script that crashed at hour three. The discovery is buried under the logistics.

What if you could hand all of that to a tireless lab assistant who takes instructions in plain language? That assistant is starting to exist. It is worth understanding how it actually works — past the headlines, and without the hype.

The leap

Describing versus doing

Most people have met AI as a chatbot: you ask, it writes back. That kind of AI describes — it can tell you, fluently, how an experiment would work. A research agent is different in one decisive way: it does the experiment.¹ It writes the code, runs it on a real computer,A "real computer"— a sandbox: an isolated machine the agent fully controls, so it can install tools and run code without touching anything it shouldn't. looks at what came out, notices it broke, fixes it, and runs it again — without waiting for you at every step.²

The gap between describing and doing is the whole story.
Everything else is detail.

The loop

Four moves a research agent makes

Strip away the jargon and a research agent runs the same four-move loop a careful scientist does — just faster, and around the clock.³

The same loop, run fast and around the clock: gather what's known, break it into steps, do the work in a sandbox, then verify — and begin again.

01 — Read

Gather what is already known

It starts where research starts: with the literature. The agent searches papers and databases, reads them, and pulls out the facts relevant to your question — with citations, so you can check its work.In SciencePal, a dedicated Survey Agent does this: multi-database literature search across sources like arXiv, PubMed, IEEE and ACM, with citation management.

02 — Plan

Break the question into steps

A big question is too big to answer in one shot, so the agent breaks it into a checklist of concrete steps. For hard problems it can split the work across specialist sub-agents — a literature reader, a coder, a reviewer — that share one workspace, like a small lab team.⁴

03 — Run

Do the work, inside a sandbox

This is the leap. Inside an isolated sandbox — a private computer it cannot break anything important with — the agent writes and runs real code, drives professional tools, browses the web, and looks at the output. When something fails, it reads the error and tries again. The work is executed, not narrated.Real tools, preloaded. SciencePal's protein agent runs tools like ChimeraX, PyMOL, BioPython and MDAnalysis; its materials agent runs Quantum ESPRESSO and LAMMPS — inside the sandbox, ready to go.

04 — Check

Verify it, and remember it

A result you cannot trust is not a result. So the agent validates what it produced and keeps a trail: which command, which paper, which line of code led to each claim, so a human can audit it. And it remembers — what it learned this week is still there next week.Memory that compounds. SciencePal keeps a persistent memory across sessions, and you can hand it your lab's own methods as reusable skills.

In the wild

A "virtual lab" that designed real antibodies

This is not hypothetical. In a 2025 study in Nature, a team of AI agents — a "principal investigator" agent directing specialists such as a chemist and a computer scientist, with a critic agent to push back — designed new nanobodies (tiny antibodies) against SARS-CoV-2 variants.⁵ The designs were not left on a screen. Experimental validation in the lab found, in the authors' words, "a range of functional nanobodies with promising binding profiles across SARS-CoV-2 variants." A human researcher set the agenda and checked the work; the agents did the legwork.

Swanson et al., Nature (2025): a principal-investigator agent directed specialist agents and a critic; the designs were validated at the bench.

The fine print

Powerful, not magic

It would be a disservice to stop at the success stories. When researchers built careful benchmarks to test AI agents on real, data-driven science tasks, the agents solved only a fraction without human help — and the gap between a slick demo and a reliable result is exactly what rigorous evaluation keeps exposing.⁶ The honest framing is a co-pilot, not an autopilot: the AI multiplies a scientist's reach, while the scientist still owns the questions, the judgment, and the responsibility for what is true.

Why it matters

Science with the busywork removed

If the plumbing really can be handed off, the consequences reach past any one lab. The months a team spends building and gluing tools could shrink to days. Methods that used to demand a specialist and a compute cluster start to come within reach of a curious student or a small clinic, because the tooling is shared and the interface is plain language.⁷ And because every step leaves a trail, the science that comes out can be easier to reproduce, not harder.

From the literature to the lab bench — one continuous, traceable workflow.

Where SciencePal fits

SciencePal is built to be this kind of research agent. It runs in your browser, with no setup, on top of an open-source engine. Behind the chat box is the loop above: a sandboxed computer where agents execute real work, a roster of domain specialists — for proteins, biology, antibodies, materials, plasma physics, clinical data, literature surveys and more — persistent memory, and the ability to teach it your own lab's methods. The aim is narrow and concrete: take the tooling off the scientist's plate, and give the judgment back.