Today marks the third week of implementing FDE for the AI town. I’ve abandoned the previous approach of conducting requirements research first, then deploying the Agent, and finally having the boss pressure employees to use it. Instead, I’ve shifted to letting employees build

智昊: Latest Frontline Report on FDE Implementation with Training-Only Approach

Latest FDE implementation report. Sharing some new ideas.

This is the third week of FDE implementation for the AI town. I’ve abandoned the previous approach of conducting requirements research first, then deploying the Agent, and finally having the boss pressure employees to use it. Instead, I’ve shifted to letting employees build, use, and debug Agents themselves.

Week 1: AI First Workflow Introduction

In the first week, I had IT replace all computers for the KUs participating in the FDE implementation. Now their computers are uniformly equipped with 32GB of memory, capable of running 5+ Agents, which is sufficient for this stage.

Then, I had them use various expert libraries built into Workbuddy to complete the following tasks:

1. Graphic design: posters, manuals, knowledge transfer (KT) boards, banners, etc.
2. Course design: AI-enhanced workflows for the research center courses. This is mostly done by me now because I also need to output courses for external use. The basic workflow is now running.
3. Knowledge base: Everyone created their own knowledge base, and we’re now in the dogfooding trial phase. The design manager’s setup is currently running the most stably, with strong AI recall capabilities. The knowledge bases I helped them build aren’t RAG-based but index-powered grep systems—extremely powerful. Image extraction adaptation rate is 100%, but building the indexes was quite tedious. Of course, the effort was on their part, not mine.

These tasks weren’t explicitly assigned by me. They were developed through our on-site discussions after I showed them how I use the tools, followed by independent development on their part, and finally reported back to me.

That’s the first week. My role was primarily training and guidance. I helped install two custom-built workstations and conducted three sessions, each under two hours.

I’ll reiterate: production requirements for AI and AI bloggers’ expectations for AI are completely different things. The AI capability level needed for production currently is fully met by a middle school student’s level—there’s no need for a doctoral level.

Week 2: From Workflows to Agents

By the second week, they had thoroughly used all relevant Workbuddy expert libraries. I then started guiding them to build custom Agents. Prior to this, I created two Agents: a To-Do Agent and a Consultant Agent, used respectively for managing tasks and discussing Agent development. The first purpose was to provide a template, and the second was to manage the FDE participants.

Here, I recommend using Feishu (Lark). To-dos can be managed through a chat group, while the Consultant Agent can operate in a topic group. The chat group allows the Agent to update multidimensional tables in real-time, and the topic group enables the Agent to log to cloud documents, making it easy to review discussion results between employees and the Agent.

(Screenshots of To-Do Chat Group Testing)

(Screenshots of To-Do Chat Group Testing)

(Screenshots of Topic Group Discussion Testing)

(Screenshots of Topic Group Discussion Testing)

(Screenshot of Single Topic Testing)

(Screenshot of Single Topic Testing)

The specific content cannot be fully displayed here. This solution has been online and in use for two days; we’ll observe it for another week.

The choice of these two Agents is strategic. The core idea is: when employees build an Agent and complete communication with the FDE Consultant Agent, the Consultant will directly send tasks to the To-Do Manager Agent. These tasks are then assigned to specific employees, allowing me to see what Agents they plan to build. Meanwhile, the output of the Consultant Agent’s consultation is a set of prompts. Employees copy and paste these prompts into Workbuddy, which then directly helps them build the Agent.

I’ve organized and prepared all corresponding Skills on their local machines. Of course, Workbuddy’s effectiveness will depend on how well they build things this week. If Workbuddy doesn’t perform well, I’ll switch entirely to Thincoder, which has been my primary coding Agent recently.

The overall concept is to have the FDE Consultant Agent guide employees to identify pain points and solutions within their current business workflows, then directly convert these into prompts. Employees then use these prompts with their local Coding Agents to complete the overall build and deployment.

All I need to do is sip tea and then review the results.

Week 3: Agents & Internal Tool Sharing Platform

This week, I plan to set up their company’s internal website platform. The prototype is already running on my local machine. Next, after they finish building their Agents and internal tools, they’ll conduct weekly review meetings. Once approved, they’ll use their local Agents to package and upload everything to the company server. Then everyone within the entire local network can use it.

The Real Purpose of Writing This Article

The servers are still being procured. If anyone sells company intranet servers, please contact me. I don’t need a compact case server—just a micro-server will suffice. The total user base is only about 20 people; we don’t require particularly high concurrency.

Cover prompt source: Adrian Punk

https://x.com/AdrianPunk115/status/2098366626571989034