Skip to content

How agentic AI is affecting researchers' work

Overview

The last lesson made a broad argument: AI is everywhere, people use it for surprising things, and it quietly costs something even when it works.

This lesson gets specific. What does that look like inside a research group: in the daily workflow of a PhD student, a postdoc, a senior researcher?

What "agentic" means in research

Most people picture AI at work as a chatbot: you ask, it answers, you paste the answer into a document, but this is not what is meant by an agentic system.

An agentic system takes a goal, breaks it into steps, uses tools, runs code for instance, reads the output, decides what to try next, and iterates until it produces something that looks finished, importantly, it can do this without being prompted step by step. In research, that means AI is no longer sitting beside your workflow; it's participating in it. Some people in our group already work this way every day.

What our department told us

We ran an informal survey of about 22 colleagues — PhD students, postdocs, researchers, senior scientists. Not a representative sample, but enough to sketch a picture.

Three findings stood out:

People use AI for more than half their tasks. Not one tool for one job, but continuous switching to it through the day. AI is used mostly for coding, data preprocessing, writing, and reading papers.

What people enjoy, and what they don't, look completely different. The upside answers were things like "my brain has an extended memory now" and "new topics are easier to get familiar with." The downside answers told a different story: "I feel that I often ask AI instead of thinking." "Overreliance has caused me to forget some basic fundamentals, and I don't feel like I truly learn how to code with it." A postdoc noted AI "sneaks in silent bugs that are very hard to catch." A senior researcher said: "I now spend most of my time working on 'boring' admin-style tasks, or checking code written by an LLM instead of doing any programming myself."

Concerns split by career stage. PhD students worried about career prospects, skill development, and keeping pace. Senior researchers worried more about how to evaluate contributions and judge AI's answer, but were, overall, more positive about AI than the students.

Three colleagues, three experiences

Recall the three colleagues from the last lesson.

The first had stopped enjoying coding and debugging: the creative part of the work. Now he mostly asks AI and moves on once it works. He hasn't lost the skill, just the pleasure in it. That's a real cost, one people often only admitted in a wider group, because they felt they were supposed to be happy about being more productive.

The second felt the opposite. AI cleared away the parts of the job he'd always disliked, leaving more room for the parts he loved, and more time to think before committing to an idea.

A third pattern came up repeatedly among PhD students: not being sure whether they were becoming a better researcher or just a faster one. Raising this with senior colleagues actually relieved a lot of people: the consensus was that the point of training is to build judgement and the ability to identify a good question.

How to balance all this is still an open question.

The distinction that matters

Agentic AI can drive a whole research workflow: reading files, running experiments, editing code, generating figures, reporting results. None of that means it understands what any of it is for or if it really understands what it is supposed to do. Skimming the code it produced isn't enough; you still have to ask whether the question was worth asking, and whether the answer means what it appears to mean.

Where this leaves us

Agentic AI is genuinely useful in research, and the productivity gains our colleagues reported are real. At the same time, it's doing something to us worth examining carefully.

The next lesson gets concrete about which skills are becoming more valuable in an AI-rich research environment.

If you take one thing from this: an agent can act on your behalf, but only you can understand.

Also see