跪拜 Guibai
← All articles
Agent · MCP · Artificial Intelligence

Building a Digital Human Agent for R&D: A Three-Layer Architecture with Offline Knowledge, MCP, and Personalized Skills

By vivo互联网技术 ·
Read original on juejin.cn ↗ Google Translate ↗ Alt translation

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.

Summary

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.

Takeaways
— The system is built on OpenTask, a general-purpose agent framework similar to WorkBuddy/CodeX.
— The knowledge base is fully offline and private, organized into project private library, personal repository, and team offline document library.
— MCP provides real-time service info, RAG topology, node details, source code retrieval, online service requests, and online document pulls.
— MCP consumers must implement timeout circuit breaking and fallback to offline knowledge on failure.
— Skills are triggered implicitly by keywords/semantics or explicitly via commands like /help or @code-review.
— Skill and Prompt are deeply coupled: Skill provides the flow skeleton, Prompt provides the brain logic for each step.
— Output style can switch between technical and popularized via persona or Skill+Prompt differentiation.
— In a real example, the agent correctly identified that a price filter could not distinguish 5000-6000 from 6000+ because the classification enum only had six hardcoded ranges.
— The agent can generate scaffold code, unit tests, Dockerfiles, and CI/CD scripts from a requirement document.
— The system has replaced over 80% of tedious tasks like problem lookup, troubleshooting, and data verification.
— Future plans include automatic incremental knowledge base updates, Skill decomposition into atomic skills, loop closed-loop monitoring, and few-shot autonomous learning.
Conclusions

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.

Concepts & terms
MCP
Model Context Protocol, a standardized interface that lets an agent call external tools and APIs. In this architecture, MCP acts as a uniform gateway to online business systems, returning real-time data like service info, node status, and source code.
Skill
A reusable, customizable task skill that encodes an engineer's personal troubleshooting logic, workflow, and habits. Skills orchestrate the large model, offline knowledge, and MCP calls according to a fixed procedure.
RAG topology
The dependency graph of services or processing nodes in a retrieval-augmented generation system, showing how data flows between components.
OpenTask
A general-purpose intelligent agent framework (similar to WorkBuddy/CodeX) that provides the core reasoning engine with long-context support and offline inference, used here as the foundation for building the digital human agent.
Source: juejin.cn ↗ Google Translate ↗ Backup ↗