Tencent's Digital Employees Achieve 10x Speedup: Rewritten Agent Architecture Delivers Minutes-Level Data Analysis

Traditional data analysis workflows require 1-2 days for an operational weekly report. Tencent's PCG Big Data Platform developed AI-powered digital employees that complete the same tasks in minutes, achieving a 10x speedup.

The core insight is that generic large models fail in data analysis scenarios: they lack business terminology understanding, produce SQL errors, suffer from context explosion, and cannot execute full pipelines. The bottleneck is not SQL generation ability but the model's inability to grasp business context, complete end-to-end workflows, and retain reusable experience.

Tencent's solution follows the formula: Good Agent = Good Model × Good Architecture × Good Knowledge .

Model Layer: LLM scheduling and context compression ensure stable model operation.

Architecture Layer: Agent core plus tools and protocols enable models to combine capabilities.

Knowledge Layer: Data semantic library, Skill system, and self-evolution make models understand business; memory system and context engineering deepen experience with use.

This modular architecture addresses eight deployment challenges: knowledge fragmentation, workflow constraints, context dilemma, memory loss, single-model limits, unreliable tool calls, process reuse difficulty, and experience non-accumulation. Each challenge maps to a specific architectural module.

The digital employee is not merely a larger Copilot but an end-to-end executor that independently receives requirements, plans analysis paths, invokes tools, and generates deliverables. It handles operational weekly reports, generates conversational HTML dashboards, and runs scheduled daily/weekly reports — replacing multiple manual steps and cutting task time from days to minutes.

Human-AI collaboration boundaries are explicitly designed: intelligent querying is fully automated; deep analysis reports require human confirmation at key nodes; compliance and high-risk actions remain human-led. This "full automation where appropriate, human fallback where needed" boundary design is the key engineering experience for enterprise adoption.

Evaluation-Driven Development (EDD) quantifies success across three dimensions: end-to-end task completion rate, delivery quality, and human intervention rate.

The architecture is packaged as a reusable modular template with eight components: data semantic layer, Skill system, memory system, context engineering, LLM scheduling, tools and protocols, skill system, and self-evolution.

This work was presented at DACon 2026 Beijing by Xie Yuan-zhen, Senior Algorithm Researcher at Tencent PCG Big Data Platform, who has published 11 papers at venues including ACL, ICDM, KBS, WSDM, and CIKM.