Before code review even starts, agents let us build custom tooling and formal models. Here’s how six months of building with AI agents helped us find real issues in our Miden zkVM audit.