AI-Generated Hardware Designs Have to Work, Not Just Get Ordered

AI-generated hardware designs are only useful if they work when they are manufactured. Coding agents like Claude Code and Codex can now generate schematics, PCB layouts, and CAD models. But generating a design and manufacturing something that actually works are two very different things.

There’s a big problem in hardware right now that nobody is talking about.

Manufacturers are already dealing with AI-generated designs that don’t work. The result is unhappy customers, refunds, and costly disputes. Some manufacturers are actively trying to stop agents from placing orders.

“Supply chain is a search problem”

Last week, Dylan Morris wrote: “Supply chain is a search problem.” Caleb from OSH Cut agreed, and added that agent-friendly manufacturers will soak up demand. When every modern shop offers its capabilities, quotes, and DFM through an API, search gets easy.

I agree with both of them. Manufacturers with good APIs will win. Agents will order a lot of hardware.

But I think we disagree on where the problem is. An agent that finds the right shop in two seconds still sends that shop a design. If the design doesn’t work, faster search only gets you a bad board, or a bent part that doesn’t fit, faster.

“Making it easier for agents to order hardware doesn’t solve the problem. Making sure the hardware actually works does!” – Matthias Wagner, Co-founder

Why AI-generated designs fail at the factory

A coding agent is great at writing files. A KiCad schematic, a STEP model, and a Gerber set are all files. So the agent writes them, and they look right.

The physical world doesn’t care how the files look. It cares about things like these:

• A regulator that is out of stock everywhere, or that went end-of-life last year.
• A footprint that doesn’t match the real package, so the part can’t be assembled.
• A trace, a drill, or a clearance that your board house can’t build.
• An LDO that gets far too hot at the real load current.
• A USB-C connector that sits 2 mm away from the opening in the enclosure.
• A sheet-metal flange that is too short for the bend, or a hole too close to the bend line.
• A bracket that cracks at the first drop, or an enclosure with no path for the heat.

None of these show up as an error in a text file. They show up as a refund request, three weeks later, at the manufacturer. Then everybody loses: the customer, the shop, and trust in AI-designed hardware.

Working hardware, not CAD files

At Flux, our focus has always been on delivering working, manufacturable hardware, not just generating CAD files!

That’s why the Flux agent doesn’t only write the design. It checks the design against the physical world before anything gets ordered. The electronics and the mechanical parts of the device are in one project, so the checks cover the complete device:

• Parts you can buy. Flux checks live stock and pricing at distributors such as Digi-Key, Mouser, LCSC, and Arrow. It reads the datasheet of every part and checks the design against its pinouts, ratings, and limits.
• Rules your manufacturer can build. You set your layout rules and your manufacturer’s capabilities once. Electrical and manufacturing rule checks run on every schematic and board.
• Parts that fit together. The board, the enclosure, the brackets, and the wire harnesses are in the same project. Flux checks fit and clearance on the assembled device. A connector that moves on the board gets checked against the opening in the enclosure.
• Simulation before the build. Flux runs SPICE simulation of circuits, full-wave electromagnetic simulation of the routed board, thermal and airflow simulation of the board and the enclosure, and stress analysis of mechanical parts. The agent reads the results and fixes the design.
• Your team’s experience. Preferred parts, design guidelines, and lessons from the last product become team knowledge. Every project follows them.

Then the agent gets the quotes for the board, the enclosure, and the other parts, and places the orders. By the time a design gets to a manufacturer, it was already checked against the parts market, the manufacturer’s rules, the enclosure, and simulation. I wrote more about that last step in Design Is the Easy Part.

Do we always succeed? No. But we’ve invested years of engineering into the tools, the context, and the validation needed to make that happen.

General-purpose agent companies won’t build this

General-purpose agent companies will never make that same investment. They’re focused on code, as they should be.

Hardware engineering needs an entirely different stack of specialized tools and domain expertise: live part data, manufacturer rules, PCB and mechanical CAD kernels, circuit, EM, thermal and stress simulation, and a lot of knowledge about what fails on the bench. That stack takes years to build. We started in 2019.

You don’t have to choose, either. Use OpenAI, Anthropic, or Gemini models inside Flux, and they get the hardware tools and the checks. Or connect your existing agent to Flux over MCP. Either way, the design gets checked before it gets built.

Where this ends

Agents will order hardware. A lot of it. The manufacturers who let them will soak up demand, just like Caleb says.

But the agents that win won’t be the ones that order fastest. They’ll be the ones whose orders work. When the first article comes off the line, the board boots, the connector lines up with the enclosure, and nobody asks for a refund.

That’s the bar. Not a design. Not an order. Working hardware!

FAQs

Can coding agents like Claude Code and Codex design hardware?

Why do AI-generated PCB designs fail at manufacturing?

How does Flux check that an AI-generated design works?

Does Flux check mechanical parts as well as the PCB?

Can I use my own AI agent with Flux?

Describe the device you need. Flux designs the board and the mechanical parts, checks them, gets the quotes, and orders them. Start Free Trial.