ABSTRACT

Agents can now carry an aerospace conceptual design through the coupled loop of performance, mass, power, thermal, and cost in hours rather than review cycles. This talk presents our work on agent harnesses that close the design-simulate-evaluate loop automatically, using simulation as a verifiable reward signal to keep agents grounded rather than hallucinating plausible-but-broken designs.

 

In one multidisciplinary study, the first complete trade study came sixteen minutes after the acceptance plan was committed, and the design went through five full report versions in a day. That study ended by listing what had to be measured before its claim could close: material properties, noise, pointing jitter, calibration drift, and supplier quotes for cost.

 

How do we actually get to a design and verification harness? We will explore what industry is doing and how we validate with others. Once analysis is cheap, those items set the schedule. The binding constraints become hardware evidence, materials and dynamics testing, credible cost bases, and production convergence. We'll discuss the architecture, the failure modes we've hit, and the speedups we're seeing across hardware and software workflows.

 

The rest of the talk is a proposal, not a demonstrated result. We argue that agentic engineering should move into the physical world. Test plans should come from the same acceptance check that judged the design, so each test is chosen for the uncertainty it retires. Measured samples should feed back into the model as they arrive. Testability, manufacturability, and supplier signals should enter the trade before the downselect, while they can still change the answer.