Generative design had a quiet failure that the industry has mostly stopped talking about. It produced two hundred massing options in an afternoon, and then a design team had to pick one, with no basis for choosing beyond geometry and instinct. Producing options was never the constraint. Evaluating them was.
Predictive design inverts the emphasis. Fewer options, each one carrying a forecast of how it behaves once built.
Four questions per option
When an architect tests a facade alternative, four separate teams eventually answer four separate questions about it, usually weeks apart and often after the option has already been committed. What changes with an agentic approach is that all four answers arrive at the moment the option is drawn.
- Structural: what does this do to the frame, the spans, and the foundation loads, and does it push any element into a different section size.
- Thermal: what is the resulting U-value, where does the assembly bridge, and does it hold against the Part L backstop with the current TGD in force.
- Carbon: what is the embodied carbon of the buildup per square metre, and how does that trade against the operational saving over a sixty year study period.
- Cost: what is the capital cost delta, and what happens to it under the procurement route actually being used.
The spread between the top and bottom option is 217 kgCO₂e per square metre. On a fifteen thousand square metre facade area that is a difference of roughly three thousand tonnes, which is the sort of number that decides whether a scheme meets a client sustainability commitment or quietly abandons it at stage four.
Why this needs agents rather than a plugin
Every one of those four analyses already exists as software. Structural packages, thermal modelling tools, carbon calculators and cost planning systems have been around for decades. The reason they are not used at option stage is not capability. It is friction. Each one requires a different model export, a different set of assumptions, a different specialist, and a turnaround measured in days.
An agent absorbs that friction. It reads the option out of the model, prepares the input each analysis needs, runs them, and reconciles the results into a single comparison. The architect keeps designing.
The design does not get made by the agent. The consequences of the design arrive early enough to still be design decisions.
Risk prediction against project history
The second capability is less mature and more interesting. A practice that has delivered two hundred projects holds, in its archive, a record of every place a design ran into trouble. Where clashes concentrated. Which details attracted the most technical queries. Which assemblies drew planning conditions.
That archive is almost never used, because reading it is nobody's job. An agent that scores an evolving design against it produces something no individual can: an early warning that this particular junction, at this particular scale, has caused constructability problems on four previous schemes.
The escalation curve is the entire commercial argument for predictive design. Nothing about it is new. What is new is having a system that can act on it without needing a person to remember to look.
Starting position
Practices that get value from this start narrow. One building type, one assembly family, one metric. Embodied carbon is usually the right first target, because the calculation is well defined, the data is available, and there is external pressure making the answer commercially relevant rather than academic.