One question I keep coming back to is how we will interact with computers in the future, especially for experimental work. I work for a relatively small company and right now I am the only person in my department. That makes it easy to try things. So I changed my whole workflow to one that is AI first, meaning I mostly control my computer through voice now. I dictate what I want to Claude Code, which has access to the folder that holds everything about how we work. The raw materials I use, past projects, general learnings and good practices, and the test methods we have developed. That is all the context anyone would need to work in our lab, and it is enough for Claude to do the work I hand it. It sets up a new project, which means creating the recipes and preparing the documentation for the test results, up to the point where I only need to go to the lab and run the experiments. When the experiments are done it makes the plots and writes the report.

This works quite well, but in my opinion not yet well enough. The Pareto principle applies very well here. Through this new AI first workflow you get 80 percent of the work done really fast. Probably in even less than 20 percent of the time. However the last 20 percent is sometimes a little bit slow and frustrating. Because it’s always little things that would be really easy to change in the “normal workflow”. Like changing the factor level from 5 percent to 3 percent. Or adding and removing factors when performing ANOVA. Or just exploring the data, which is much easier in an interface where you can click around.

So I think the future might be a combination of both worlds. Something that’s known as generative UI. You use your voice to describe the problem or goal you have. Based on this description, an AI agent creates a custom UI only for you and your specific project that only contains what you need at the moment. You won’t be distracted by too many choices. Like when you are facing a screening problem, you don’t have the option to choose a response surface design because at the moment it’s not relevant to you. Only when you need it are you presented with this option. Or when you ask for it of course.

Let’s take some time to look at how this could work, using the same example from the last essay.

Setting the levels

In the last essay the agent proposed the factor ranges as a table that I was able to change but only by prompting again.

Claude
Below is a table with one row per factor. Fill in the ranges you want to test. I already added typical values to start from, but change anything you want.
FactorLow settingHigh setting
Curing temperature60 °C80 °C
Hardener typeType AType B
Catalyst level1 %5 %
Curing time30 min60 min
You
5 % catalyst is too much. Reduce that to 3 %. The rest is fine.

This is quite tedious and it would have been much easier if I could have just changed it myself. With a generated interface the same change is immediate. It’s almost like any DoE software today, but much simpler in its layout, with fewer menus and less clutter to confuse you.

claude assembled for: coating hardness study
Claude
Here are the factor ranges. I filled in typical values. Adjust anything, then I'll build the design.
Curing temperature °C
Hardener type Type A/Type B
Catalyst level %
Curing time min
Set the catalyst high level to 3 %. The design updates as you type.
RunTempHardenerCatalystTime

Analyzing the results

When I try to make sense of the results, I find it really helpful to just click around and look at different plots to get an idea about the data. That is tedious when all you have is a chat interface, because it turns into a lot of back and forth.

You
Please show me the main effect plots.
Claude
Here you go. Temperature has the biggest effect, then curing time and catalyst concentration. The hardener has almost no effect on the hardness.
curing temperaturecuring timecatalysthardener type+19.0+16.5+14.0+1.5effect on hardness, oscillations →
Average effect of each factor on the hardness of the coating.
You
What about interactions?
Claude
There is a relatively strong interaction between catalyst concentration and curing temperature. At high temperatures the catalyst has almost no effect. At low temperatures that effect is large. Do you want to see the interaction plot as well?
You
Yes please.
Claude
Here you go.
hardness, oscillations →4582.5≈ 7860 °C80 °C+32 oscillations1 %3 %catalyst concentration
Curing temperature and catalyst interact. At 60 °C more catalyst helps a lot, at 80 °C it barely matters.
You
Show me the other interaction plots as well please.
Claude
Of course.
temperature × time60 °C80 °C30 min60 mincatalyst × time30 min60 min1 %3 %both sets of lines run parallel, so these factors do not interact
The remaining interactions are flat. Those factors add up on their own and do not interact.

With an interface generated just for my experiment, the plots are already there. It feels almost like a conventional DoE software package, but with far fewer menus and distractions. I move around on my own. The assistant gives me a first evaluation of the results, but I can look through all the plots myself and even extend the interface if I want, like an ANOVA or planning the next runs for my project. I just need to ask and a new tab appears. Try it, the widget below is interactive.

claude 8 runs · hardness
Claude
Results are in. Curing temperature, time and catalyst drive the hardness. Hardener does nothing, so I set it aside. The interaction that matters is temperature × catalyst. It's on the second tab.
Ask
curing temperature+19.0 curing time+16.5 catalyst+14.0 hardener type+1.5 effect on hardness, oscillations →

Hover a bar for its value. Hardener barely moves the hardness, so it is set aside.

This is already possible today

What we just saw is not really a glimpse of the future. You can set this up today with Claude Code. I also imagine an autonomy slider, where you decide how much freedom to give the agent. There could even be a learning mode for people who are new to this.

The software finally adapts to you and your needs, not the other way around.