Building a New AI‑Driven Project Management Paradigm: The Redbook PMO’s Agentic Journey

Over the past year the Xiaohongshu PMO team performed four major iterations of an AI project‑management Agent, each tightly coupled with advances in large‑model and Agent engineering.

Version 1.0 – AI Project Consultant : The Agent acted only as a knowledge‑base Q&A system, answering basic project‑management questions using a RAG pipeline (intent detection → knowledge base → LLM generation). The main output was a test set of 10+ representative cases, which served as the Agent’s unit tests.

Version 2.0 – Agent + Multi‑Channel Execution : The design introduced a “Agent + Sub‑Agent” principle, enabling the Agent to act beyond answering questions and to perform actions such as Todo management, progress nudging, weekly‑report summarisation, and automatic notifications. Interaction moved to the internal IM platform, and a unified notification interface and a master‑sub‑Agent JSON protocol were added.

Version 3.0 – Personal Assistant on OpenClaw : By adopting the OpenClaw framework, the team built a personal assistant with long‑memory capabilities. Three core upgrades were made:

UserProfile : four‑dimensional user portrait (preferences, behaviours, FAQs, related projects).

SessionMessage : three‑tier cache of dialogue history (DB → application → Agent).

KnowledgeItem : five‑type knowledge store supporting full‑text and tag search.

ContextBuilder : dynamic context assembly with automatic compression.

These components allowed the Agent to retain information across sessions and channels, turning it into a “personal project‑management assistant”. The assistant was packaged as a Skill in the internal Skill Hub and integrated with a self‑built project‑registration platform that served as the single source of truth for project data.

Version 4.0 – PMOBP Agent (Shared Project‑Group Assistant) : The Session model was upgraded from “1 person × 1 session” to “1 project × N people × M sessions”, providing a shared, dynamic context for every project group. Two new capabilities were added:

Project‑Context Routing : automatically identifies which project the user is asking about.

Role Awareness : adjusts answer granularity for managers, project leads, and team members.

The resulting architecture re‑uses OpenClaw’s gateway, session model, multi‑agent orchestration, IM plugin, Skill system, and stability mechanisms, while adding project‑specific routing and role‑sensing layers.

Design Principles Consolidated (kept from 2.0 onward):

Atomic Agents must be self‑closed – they either return a successful result or a clear failure without further master‑Agent intervention.

Composite Agents are built by composing atomic agents; nesting composite agents is prohibited to avoid dead‑loops and excessive compute consumption.

Seven Practical Lessons learned from the whole effort:

Do not wait for all infrastructure to be ready; ship with the fastest‑to‑market platform and let architecture evolve with scenarios.

The evaluation set is the Agent’s unit test; it is more valuable than the raw knowledge base.

Atomic Agents must be able to close the loop themselves.

Structured long‑memory (UserProfile, SessionMessage, KnowledgeItem, ContextBuilder) is essential; simple chat‑log storage is insufficient.

Domain Agents should act as the “Source of Truth” rather than reinventing functionality.

Project master data must precede scenario expansion – richer anchor data yields smarter agents.

In the AI era every PMO becomes a Builder, creating AI‑native products instead of merely guarding processes.

Future Outlook : The team envisions moving from “people → people + AI → AI + human assistance → AI fully in charge” and aims to embed the PMOBP Agent into every project group, making “Agentic project management” a reality across the organisation.