3 Agents. 3 LLMs. 1 Aging GPU: Engineering Parallel Inference on Bare Metal

Running multiple AI agents on a single, older GPU often results in crashes because the first model launched greedily reserves most of the available VRAM for its KV cache. A specialized C++ daemon solves this by acting as a central "bookkeeper," managing all agent requests to prevent the GPU's memory from being overcommitted. This system works by first checking if a new model will fit within a VRAM budget and only then loading it, a "book before you build" strategy that prevents out-of-memory errors. The architecture further optimizes resources by initializing the backend once and loading each unique model a single time, efficiently sharing it across any agents that need it.