Hosting Agentic Workflow Agents and Open-Source LLMs

Organizations are increasingly looking to self-host advanced Large Language Models (LLMs) to power autonomous AI agents (“agentic” systems) while maintaining control over data and infrastructure. This white paper provides a technical blueprint for hosting open-source LLMs – with a focus on Meta’s LLaMA family and the DeepSeek R1 reasoning model – in on-premises, cloud, and hybrid environments. We discuss system requirements (e.g. GPU, memory, and compute needs) for deploying models ranging from 7B to 70B+ parameters, and outline architecture patterns for different deployments. Key topics include infrastructure setup, containerization and orchestration, security considerations for sensitive data, resource optimizations (like quantization and model distillation), and monitoring strategies to manage LLM-powered agents at scale. Solution architects and AI engineers will gain practical guidance – including checklists, code snippets, and best practices – to implement reliable, efficient agentic workflows using self-hosted LLMs. Authoritative references and real-world examples are cited to reinforce recommendations.