Generative design that ships starts with constraints, not prompts.

Generative design has a portfolio problem: the internet is full of organic lattice sculptures that won 3D-printing beauty contests and never touched a production line. The technology isn't the issue. The workflow is. Generation without encoded constraints produces beautiful answers to the wrong question: geometry that ignores the mould, the budget, the assembly line and the service technician. Tools that turn a prompt into a plausible-looking part make that problem bigger, not smaller. Plausible isn't the same as producible.

Constraints first

The real work happens before any geometry is generated: encoding the package envelope, the load cases, the manufacturing process with its actual rules (minimum wall, draft, tool access), the cost model, the interfaces that cannot move. That encoding is engineering judgement made executable, and it's where most of the value lives. Get it right and the search space contains only buildable answers. Get it wrong and you've automated the production of impossibilities. We put the constraint model in front of the client's engineers for sign-off before the first run; it's their knowledge, formalised.

Search wide, verify hard

With constraints encoded, machines do what humans can't: explore thousands of architecture and geometry variants without fatigue or attachment. But breadth is only half the loop: every promising candidate goes through simulation against the physics and a manufacturability check before a human ever sees it. What reaches the review isn't "what the algorithm made"; it's the verified best of a space no manual process could have covered.

Producible or it doesn't count

Our rule for generative work is blunt: if it can't be made by the process you'll actually run, it isn't a result. It's concept art. Optimised components are re-detailed for the real process with our engineering practice, and the measured delta (mass, stiffness, thermal margin) is reported against the baseline honestly, estimates labelled as estimates.

The agent-native lens: constraints that travel with the part

When the buyer sends an assistant. An assistant shortlisting components for an engineer gains nothing from knowing an algorithm drew the part. What it can check is whether the part fits and what it does: the envelope, the mounting points and mating features, the material, the load it's rated for. Those are the constraints you encoded before generating anything, so publish them from that same model, as dimensioned data and downloadable geometry rather than a hero image of a lattice. If the listing says lighter or stiffer, say compared with what, and whether the figure was measured or estimated. "AI-designed" gives a machine nothing to verify. A stated baseline does.

When the business runs on agents. More and more of the generative pipeline is work agents can run: setting up the study from the constraint model, launching the variants, discarding those that break the process rules, ranking the rest. That makes the constraint model the most important file in the program. Keep it versioned, with an owner for every rule, so that when a supplier changes its minimum wall or the cost model moves, an agent can re-run the affected studies and list each released part that was generated against the superseded rule. The agent never sends geometry to tooling. An engineer picks the candidate, a person answers the supplier's manufacturability review, and the release records which constraint version and which inputs produced the part.

People decide

Nothing ships because an algorithm scored it highest. The final call weighs things no objective function holds (brand, service reality, assembly ergonomics, risk appetite) and belongs to people. The machines widen the search; the judgement stays human. As generation gets cheaper and more of the pipeline runs itself, that line matters more, not less: the easier options are to make, the more the value sits in the truth they were made against and the person who chose between them. That division of labour is what "AI-augmented design" means when it's a practice rather than a press release.