I've been gearing up to submit agentcad to cadbench which is an eval set for CAD models.
As I've been going through that, I've found some error scenarios where everything "looks" right to me and to the agent but is actually off by a small amount here or is missing a specific shape over there.
agentcad
Curious whether the harder problem has been generating CAD designs or giving agents enough understanding of engineering constraints to avoid bad designs in the first place.
Which side has surprised you more?
agentcad
@zaid_mallik1Β the agents will always generate something so producing quality is definitely the bigger challenge.
@jdillaΒ Makes sense. The challenge is shifting from generation to validation.
We're building something in the AI workflow space and have noticed a similar pattern β getting an output is rarely the bottleneck anymore, trusting it is. Planning to launch soon and excited to see where agentcad goes.