Building a Digital Human Agent for R&D: A Three-Layer Architecture with Offline Knowledge, MCP, and Personalized Skills
For developers drowning in repetitive debugging and context-switching, this is a concrete blueprint for encoding personal knowledge and workflow into an agent that runs unattended. The architecture shows how to combine offline privacy with online real-time data, and the 80% workload reduction figure suggests the approach has real traction inside a major tech company.
A team at vivo has built a "digital human agent" for R&D engineers, targeting the overhead of repetitive tasks like problem troubleshooting, log analysis, and data verification. The system runs on OpenTask, a general-purpose agent framework, and layers three components: a fully offline local knowledge base that stores the engineer's private project docs, troubleshooting notes, and service architecture; MCP interfaces that pull real-time data from online systems; and customizable Skills that replicate the engineer's personal reasoning habits and workflow.
The agent handles the full loop: it parses a question, retrieves relevant offline knowledge, validates against live MCP data, applies a Skill-based reasoning chain, and outputs a structured solution. It also generates standardized documents like modification suggestions and high-level design docs. The team reports that it already handles 80%+ of their tedious tasks, with future plans for loop monitoring, autonomous learning, and exposing the capability through WeChat official accounts while keeping strict read-only boundaries.
The biggest differentiator is the emphasis on personalized Skills and private knowledge, moving agents beyond generic code completion and into replicating an individual engineer's judgment.
The architecture treats MCP as the bridge between static offline data and live online systems, but the strict requirements for MCP providers (unified interface, real-time data, link for large content) reveal how fragile such integrations can be without discipline.
The 80% replacement claim is striking but the article shows a single real-world example (the price filter bug) and does not quantify the time saved or mention any limitations or failures.
The planned Loop closed-loop monitoring and autonomous learning suggest the authors recognize that one-shot question answering is not enough; true automation needs verification and recovery.
The read-only boundary for external access reflects a pragmatic approach to security: the agent can analyze and propose, but not modify systems without human approval.