Architecting autonomy: A design pattern for high-compute, governed agents in biotech
Google Cloud proposes a three-layer architecture to run compute-intensive biotech simulations while keeping proprietary data secure and workflows reliable.
Most AI agents today handle quick, stateless tasks, but biotech workflows like protein folding or compound screening require sustained, GPU-bound simulations that run for minutes or hours. Standard LLM endpoints cannot coordinate these workflows without triggering timeouts, context overflows, or exposing sensitive chemical data to external networks. To address this, a three-layer pattern separates cognitive orchestration from heavy computation and embeds both within an enterprise governance plane tailored for drug discovery. The design aims to scale autonomous agents while maintaining control over data and execution.
The execution layer runs scientific workloads such as ESMFold for protein folding and GenMol for generative design as NVIDIA Inference Microservices on Google Kubernetes Engine using G4 VMs with NVIDIA RTX PRO 6000 Blackwell GPUs. To avoid cold-start delays, the architecture maintains warm pools of GKE nodes with preloaded model weights in GPU memory, ensuring containers are ready for inference. GKE also handles pod lifecycles and restarts without disrupting the underlying VM infrastructure, reducing the need for custom CUDA tuning and lowering engineering overhead.
Proprietary chemical structures and IP must remain within company custody, so the architecture isolates execution in a private VPC. Researchers can initially query NVIDIA-hosted endpoints during exploration, then deploy the same containerized NIMs internally when transitioning to lead candidates. The orchestration logic remains unchanged; only the target URI shifts from an external proxy to an internal endpoint, ensuring no structural data leaves the enterprise perimeter.
The cognitive layer, built on Agent Runtime within the Gemini platform, manages workflow sequencing using a hierarchical multi-agent structure. For example, when a user submits a task such as analyzing a candidate against a target, the system applies orchestration patterns that evaluate candidates against dynamic, target-specific profiles. The design emphasizes optimizing I/O and execution topology to prevent pipeline bottlenecks, such as reducing payload sizes, aligning concurrency with backend queues, and passing heavy payloads by reference to avoid token bloat or validation failures.