WorkBuddy Practical: 24 - Multi-Agent Architecture and Collaboration

Is the task worth splitting?

The more of the following conditions apply, the better suited the task is to multiple Agents:

• At least two subtasks can proceed independently;
• The subtasks require different methods, materials, or tools;
• The handoff format for each output can be defined clearly;
• Parallel execution can materially shorten waiting time;
• There is a clear overall owner and final acceptance check;
• The budget allows multiple rounds of calls.

Writing one email, summarizing one PDF, or formatting one spreadsheet does not require an expert team.

Case: a product-promotion video expert team

Background

HyperFrames is HeyGen's open-source video-rendering framework (17.7K GitHub stars at the time of the case). Its main characteristic is that it is friendly to AI Agents: an Agent can automatically generate HTML-based video frames and render the output. Product-promotion videos also follow a relatively stable pattern—no voice-over or actors are required; the main elements are product demonstrations, concept captions, and background music. This makes the task suitable for an Agent team.

Process design

The full workflow of the product-promotion expert team is:

Role contracts

Expert-team demonstration

Before making the product video, place the relevant materials in the workspace.

I want you to make a product-promotion video for Tencent WorkBuddy's latest Expert Team, focused on OPC scenarios. I have placed some materials in the current workspace. The finished video can use an Apple-like style and real software interfaces. Make the whole process fully automatic.

The lead first receives the task and splits “make a product video” into subtasks: understand what the WorkBuddy Expert Team is, who it serves, and its core value; decide on the narrative structure, shot count, and pacing; then assign assets, editing, and music.

The Brief role researches the WorkBuddy website, product introduction, and Expert Team list, then writes a brief covering what the product is, its target users, and the strongest points to fit into 60 seconds.

The storyboard artist uses the brief to split 60 seconds into seven shots, specifying timecodes, visuals, text, transitions, motion, and required asset types for each shot.

The asset producer and editor then work on their respective tasks. One generates or captures product screenshots and concept images; the other places the assets into HyperFrames according to the storyboard and renders each shot into an MP4 clip.

The music producer does more than write a generic “tech BGM” prompt. It reads the storyboard, studies the emotion curve, marks where beats should land on a product reveal, where the track should drop for breathing room, and where a hit point should support the CTA. It then calls a music model to generate BGM candidates.

Finally, the lead integrates every output, runs the final edit, and delivers the finished video. Human involvement is limited to key decisions such as storyboard order, BGM preference, and caption changes.

<video controls preload="metadata" src="./assets/006_asset_Um2SbSClHo.mp4"></video>

Shared artifact layer

Multiple Agents should not each maintain their own copy of product facts. Create one artifact path:

project/
├── brief.md                  # Product brief, produced by Brief role and accepted by lead
├── storyboard.md             # Storyboard, produced by storyboard artist and accepted by lead
├── assets/                   # Assets produced by asset producer
│   ├── screenshots/
│   └── concepts/
├── clips/                    # Frame-by-frame clips produced by editor
├── bgm/                      # BGM candidates produced by music producer
└── output/final.mp4          # Final video integrated by lead

Downstream roles read only upstream artifacts that have been accepted. Key details should not be passed between roles through conversation.

Parallel and serial work

Can run in parallel: asset generation and editing preparation; rendering different shot segments.

Must run serially: confirm the brief before writing the storyboard; confirm the storyboard before generating assets; wait for assets before editing; complete the video before mixing the music.

A parallel plan must state where work converges. Assets and editing can be prepared in parallel, but final assembly must wait until all assets are ready.

The lead's responsibilities

The lead (producer) controls the workflow:

• Interpret the user's task and maintain subtask status;
• Distribute only the minimum required context to each role;
• Check that upstream artifacts meet the handoff format;
• Decide whether to parallelize, wait, or retry;
• Ask the user for decisions at key points, such as storyboard approval and BGM selection;
• Integrate every artifact and perform the final assembly;
• Check visual consistency, caption alignment, and BGM timing.

Three points that require human approval

1. Brief approval: Is the positioning, target user, and core selling point accurate?
2. Storyboard approval: Does the narrative structure, shot count, and pacing fit the goal?
3. BGM selection: Does the emotional style match the finished video?

Agents generate and execute; they cannot replace brand direction or aesthetic judgment.

Productization: from a self-built team to a preset expert team

Build your own team

Package the roles above as Skills and orchestrate them in an Agent framework. This suits developers who need full control over role prompts, tool permissions, and handoff formats. The requirements include defining responsibilities, designing handoff formats, and debugging parallel and serial logic.

Use a preset expert team

WorkBuddy productizes the same division of labor: the lead breaks down and assigns tasks, members execute in parallel, and the user only needs to describe the task.

To create an expert team, open Experts → My Experts → Create Expert. WorkBuddy opens a conversation and provides a format for quickly creating a custom expert.

Typical preset teams include:

Choosing between the two paths

Quality factors

Finished-video quality is mainly affected by:

• Agent base model: instruction following and reasoning directly affect storyboard quality and task decomposition;
• Image-generation model: affects screenshot clarity and concept-image quality;
• User-provided assets: placing images and videos in the workspace in advance can materially improve the video;
• Browser-tool access: an Agent that can operate a browser can capture official-site screenshots and product interfaces automatically.

Fully automatic execution suits quick videos such as open-source project introductions and product demos. For quality-sensitive work, use Agent outputs as a base and perform a second human editing pass.

Failure propagation control

A degraded delivery must state what is missing; it must not present itself as a complete result.

Multi-Agent task brief template

Goal: Make a [duration] product-promotion video for [product name].
Style: [reference style, such as Apple-like or minimalist].
Assets: [asset workspace path or provided images/videos].
Roles: lead, Brief, storyboard artist, asset producer, editor, music producer.
Approval points: user approval is required after the Brief, after the storyboard, and when selecting BGM.
Models: Agent model [specified]; image-generation model [specified].
Fully or semi-automatic: [state whether human intervention is needed at intermediate points].