We present JAM, a generative protein design system that enables fully computational design of antibodies with therapeutic-grade properties for the first time. JAM generates antibodies de novo in both single-domain (VHH) and paired (scFv/mAb) antibody formats that achieve double-digit nanomolar affinities, strong early-stage developability profiles, and precise epitope targeting without experimental optimization. We demonstrate JAM\'s capabilities across multiple therapeutic contexts, including the first fully computationally designed antibodies to multipass membrane proteins - Claudin-4 and CXCR7. Against SARS-CoV-2, JAM-designed antibodies achieved sub-nanomolar pseudovirus neutralization potency, with early stage developability metrics achieving established clinical benchmarks. We show that increasing test-time computation by allowing JAM to iteratively introspect on its outputs substantially improves both binding success rates and affinities, representing the first evidence that test-time compute scaling may extend to physical protein design systems. The entire process from design to recombinant characterization requires <6 weeks, and multiple targets can be pursued in parallel with minimal additional experimental overhead. These results establish de novo antibody design as a practical approach for therapeutic discovery, offering paths to both improved efficiency in standard workflows and new opportunities for previously intractable targets. Disclaimer: While we provide detailed descriptions of experimental methods and success metrics, we choose not disclose methodological details of JAM for commercial reasons.