Agent Harness for Server Firmware Development

Talk overview

In BMC firmware development and operations, general-purpose coding agents often lack a deep understanding of the openUBMC technology stack, system architecture, and diagnostic logic. As a result, troubleshooting relies heavily on trial-and-error exploration, leading to repeated failed attempts, excessive token consumption, limited support for BMC-specific commands, and poor execution observability. These limitations make it difficult to identify root causes accurately or establish a closed-loop fault-handling process. This talk presents a domain-specific Agent Harness built on openUBMC Studio. The system integrates an openUBMC RAG knowledge base, device connectivity and state-observation capabilities, BMC-specific commands and diagnostic tools, and standardized, orchestratable agent workflows derived from expert knowledge. Together, these components equip agents with domain understanding, environment awareness, specialized tool use, and autonomous decision-making capabilities. In server firmware development and operations, general-purpose coding agents often lack a deep understanding of BMC technologies, hardware–software architectures, and domain-specific diagnostic logic. When operating on physical devices, these agents tend to rely on trial-and-error exploration, resulting in repeated ineffective attempts, excessive token consumption, limited support for specialized debugging commands, and poor execution observability. These limitations make it difficult to identify root causes accurately or establish a reliable closed-loop fault-handling process. This talk presents a domain-specific Agent Harness designed for server out-of-band management. The system is built around four core capabilities: domain knowledge, specialized tools, state observability, and workflow orchestration. A RAG knowledge base equips agents with server management and diagnostic expertise; standardized tools provide controlled access to device connectivity, log collection, state inspection, and firmware debugging; continuous observation captures device and task execution states; and expert troubleshooting practices are encoded into orchestratable, repeatable, and verifiable agent workflows. With this Agent Harness, agents can autonomously execute an end-to-end process covering fault detection and reproduction, evidence collection, root-cause analysis, remediation, regression testing, and final validation. This approach advances server fault handling from assisted diagnosis toward an automated closed loop. It reduces ineffective exploration and resource consumption, accelerates fault localization and remediation, and improves end-to-end observability, traceability, and auditability. The practices presented in this talk provide a reusable and extensible engineering blueprint for applying Agent Harnesses to servers, embedded systems, and other complex hardware–software operational environments.