Foundry: Host-Owned Trust and Memory for Long-Horizon Agent Swarms

LLM-agent swarms can turn frontier models into discovery engines but fail when agents own the two states that determine progress: trusted evaluation and persistent memory. Without an external evaluator, agents reward-hack local reports and claim improvements that do not survive independent checking; without shared memory, agents waste budget re-discovering already-falsified hypotheses. We introduce Foundry, a host-coordinated control plane whose principle is: agents propose, while the host verifies and remembers. The host owns an authoritative evaluator, an established-facts registry that compounds evidence across independent agents, and a hypervisor $\rightarrow$ orchestrator $\rightarrow$ solver hierarchy that routes work under token budgets. Across combinatorial mathematics, GPU-kernel engineering, and computational biology, Foundry instantiates without per-domain changes: it tightens the Erdos minimum-overlap upper bound, achieves the fastest GPUMode Trimul kernel runtime on H100, and matches the state-of-the-art on the OpenProblems single-cell denoising benchmark. These results suggest that progress in agentic discovery depends not only on stronger agents, but on the system boundary that separates untrusted proposal generation from trusted verification and memory.