← ALL ENTRIES
Entry № 30 - November 16, 2026

What a Bed Taught Me About Agentic Engineering

I let an AI do the CAD design for a family bed. I gave it measurements, it drew. What that taught me about people who build software with AI without being software engineers.

I am a software engineer and a hobby woodworker. For a while now I have watched people who are not software engineers build software with AI, and I have had opinions about it. What I never had was their perspective.

Then our family needed a new bed.

Sleeping with a kid in your own bed means there is no space. So I wanted a family bed, and I wanted to build it myself. I had a few constraints: it has to stay below the light switches, it should look like it floats, and it must be 240 cm wide.

I can build that. But I have always needed a drawing first, because it makes the building so much easier. And getting to that drawing has always meant fighting with Fusion.

I had heard that the new Anthropic models are actually quite good at 3D modelling, so I let one give it a shot. Fusion has an official MCP server, so that was a match made in heaven. I gave Claude my measurements and my constraints, and it drew. I did not draw a single line in CAD.

That put me in a seat I was not used to, one I had only observed from the outside. I know what I want and I can judge the finished thing. But I did not make it myself, and I do not have deep technical expertise in the craft that produced it. I am not a technical drawer.

Step by step, not vibes

This was not vibe coding for furniture. I did not one-shot a bed and go.

We went step by step: the parameters first, then the frame, then one joint at a time, then one jig at a time. After each step I looked at the model in Fusion, turned it around, and then approved it.

I was sceptical the whole way. When something did not make sense, I asked until it did.

When the first drilling and routing jigs showed up, I wrote: "I don't understand the jigs. They're square? Shouldn't there be holes in them?" I asked why I would need two routing templates, and two became one. I asked for a picture of where a drilling guide sits and which hole comes out of it.

It also helped that this happened inside Fusion. Fusion is not a blank canvas. A parametric model either computes or it does not. Two parts either collide or they do not. The tool guarantees some consistency on its own, so I was not trusting the AI alone. I was trusting the AI inside a system with guard rails, and that felt good.

What I corrected came from my preferences and my situation

What Claude proposed first was usually reasonable. It was also generic. It was a solution for anybody, built by anybody, in any workshop.

Most of my corrections were about making it mine.

Some were taste. The first fascia boards were 40 mm thick, and that was too thick for me. Steel angles inside the frame would have been the easy joint. I find them ugly, so I took the harder joinery.

Some were my situation. Claude proposed steel drill bushings for jigs that have to survive exactly one project. We printed them instead. It planned a drilling depth that needs a longer bit than I own. It designed jigs without checking whether they print without supports on my printer.

And one I found by walking through the build in my head. There was a jig for the bolt hole in one timber, but I also have to drill into the tenon from the front, and the two holes have to line up. Nothing guided that. You only see it if you imagine yourself standing there with the drill.

This applies to agentic engineering just the same. What comes out by default is generic. You have to make it specific to the problem you want to solve, with what you have. I have written about domain thinking before, and this was that, from the other side.

Where I had nothing to fall back on

The part that took longest was a steel bracket smaller than a matchbox.

I am not a metalworker. Claude specified a stainless angle that I would cut from a bar, drill and countersink. I do not have the tool for that. But the real problem was bigger than the tool: I did not know what I would need to do, and I did not know what I could do. Is a thinner catalogue bracket strong enough? Claude said it was too weak, and later took that back. I could not judge either answer.

With wood, I can tell when a proposal is off. Here I had no knowledge of my own to fall back on, so every round was slow. I asked, got a confident answer, and had no way to check it except to ask again. In the end I settled for a bracket that fits as bought, with no metalwork at all, and I am still not sure it will fit nicely.

The same thing happens in agentic engineering. Where you know the field, the AI makes you fast. Where you do not, you cannot steer it and you cannot check it, and that is where the time goes.

What I expect to find

No wood is cut yet. Right now the design looks great.

I know that state from software. It looks great, and then I stumble over some stupid mistakes. There are always those 5 to 10 percent that the AI gets wrong. In code I find them, because I have the experience to see them. Here I will not. I am not a technical drawer, so I do not know what a wrong drawing looks like until the parts do not fit.

One of them already showed up by chance. The design called for a 250 mm wide board of Swiss stone pine. In the model that was no problem. Then I asked, almost in passing, whether a board that wide is realistic at all. It is marginal. The boards are now glued from two strips each.

The others I will find in the workshop. I will report back when there is a bed.

What this means for software engineering

This took me about two hours. Doing the design, the jigs and the plans myself would have been days and days of work.

That is the same difference non-engineers see when they build software with AI, and I understand the pull much better now. Telling them not to do it without an engineer is not going to work.

But I also see more clearly what they are missing, because I was missing it too.

They can do what I did with the wood: go step by step, look at the result, and push the generic default towards their own problem. That part works, and it works because they know their field.

They cannot do what I could not do with the steel and with the drawing itself: judge an answer in a field they have no expertise in, and spot the 5 to 10 percent that is quietly wrong.

So I think that is where the engineer's job goes. Less of the writing, more of the guard rails and the expertise behind them: a system that enforces some consistency on its own, the way Fusion did for me, and someone who knows what a wrong result looks like before it reaches production.

More entries →LinkedIn ↗
Let's build something that lasts.
engineering leadership · system architecture · or just a chat
Email me →