Foreword

In DeepSeek Harness (DSH), multi-role collaboration commonly involves a parent Agent spawning Subagents: subtasks inherit the parent’s model, tools, and ever-growing context. Even writing a commit message might go through the same costly model, with simple tasks and complex planning crammed into a single conversation, leading to high costs and context pressure.

Agent Team (limuyang2/agent-team) takes a different approach: each member is an independent root-level Agent with its own model, Session, Skills, MCP, permissions, and thinking mode. They collaborate through team tasks, messages, and a shared Workspace. Below, we introduce its positioning, capabilities, and getting started steps.

What Is This

Agent Team is a DSH workflow plugin maintained by limuyang2, with the npm package name @limuyang2/dsh-agent-team, currently at version 0.1.4, licensed under MIT. It assembles multi-Agent teams within Harness: you can mix different Providers and models, designate a single Leader, and have members work in parallel on independent conversations while sharing a common Workspace.

The plugin does not turn members into Subagents. The collaboration layer consists of team tasks, progress reports, and shared files, with members explicitly passing tasks and results without sharing an ever-expanding chat history.

Differences from Common Subagent Patterns

Dimension Common Parent Agent / Subagent Agent Team
Model Often reuses the parent model or a unified strategy Each member independently chooses Provider and model
Skills & MCP Tends to inherit a large tool catalog Each role only loads the needed Skills and MCP Servers
Context Planning, execution, and tool outputs pile up in the same session Each member has an independent Session and context window
Cost Simple tasks might still use an expensive general model Routine tasks can be assigned to smaller or specialized models
Permissions Broad permissions may spread across the entire process Each member sets default and runtime permissions independently

A software development example from the README: GPT serves as the architecture Leader responsible for planning and acceptance, a GLM coding Agent loads development Skills and MCP to modify the Workspace, and a DeepSeek Flash Commit assistant with read-only permissions reads Git Diffs to generate commit messages. The collaboration chain is: user goal → Leader assignment → coding member implementation and reporting → Leader acceptance → Commit assistant generates commit message.

Core Features

Verified from the README, the plugin currently supports the following capabilities:

  • Create reusable assistants covering roles such as planning, coding, testing, review, and documentation; they can be manually configured or generated via the built-in “Team Agent Assistant” dialog.
  • Mix Providers and models within the same team; the same assistant template can be added multiple times, each time becoming an independent member instance.
  • A full-screen workbench displays each member’s streaming output, Markdown, Think blocks, and tool calls side by side.
  • Leaders create tasks, assign members, track progress, and collect results; when policy allows, users can also communicate directly with regular members.
  • Adjust the current Session’s permission presets and thinking mode for members at runtime; view Skills loading, context usage, Token statistics, and cache hit rates.
  • Browse the shared Workspace, preview Git changes and Diffs; dynamically add or remove members, change the Leader, clear context, or disband the team.
  • A floating team entry button opens the full-screen workbench, which is draggable and locally remembered.

Regarding Skills and MCP: Agent Team uses exposed Skills and MCP Servers through Harness’s standard interfaces and is not responsible for installing or managing their lifecycle. They must first be installed into the current Profile via the corresponding Harness plugins, and then assistants can select available items from them.

Environment Requirements

  • Node.js 22.19.0+ or 24.0.0+
  • DeepSeek Harness 0.1.1-rc.2
  • Terminal available with pnpm (Harness uses it to manage Profile plugins)

If pnpm is not installed:

npm install -g pnpm

Installation and Enabling

DeepSeek Harness Web

Install the plugin into Harness’s web Profile:

npx @deepseek-ai/dsh plugin --profile web add @limuyang2/dsh-agent-team

Start Harness:

npx @deepseek-ai/dsh web

Open the address output in the terminal, usually http://127.0.0.1:3080/. After installing or replacing a plugin, Harness needs to be restarted.

DeepSeek Harness Desktop

Install the specified version into the Profile managed by Desktop:

dsh plugin add --save-exact @limuyang2/dsh-agent-team@0.1.4

After the command completes, fully exit and reopen DeepSeek Harness Desktop. --save-exact pins the Profile to 0.1.4, preventing automatic upgrades.

Uninstallation

After stopping the service by pressing Ctrl+C in the terminal running Harness:

npx @deepseek-ai/dsh plugin --profile web remove @limuyang2/dsh-agent-team

Uninstallation does not modify Harness source code or delete files in the team Workspace.

Typical Usage

1. Prepare Models in Harness

First, configure Providers, models, and credentials. Agent Team reads the model directory of the current Profile and does not store Provider API Keys.

If using GLM-5.3 and needing thinking mode, merge the following snippet into ~/.dsh/settings.yaml (replace Provider ID as actual):

llm-pi-ai:
  providers:
    zai-coding-cn:
      reasoning: high
  modelOverrides:
    glm-5.3:
      reasoningEfforts:
        off:
        minimal: minimal
        low: low
        medium: medium
        high: high
        xhigh: xhigh
        max: max
      compat:
        thinkingFormat: zai
        supportsReasoningEffort: true

After restarting Harness, you can select the level in the “Thinking Mode” of the assistant conversation toolbar.

2. Create Assistants

Go to Settings → Agent Team:

  • Start Conversation: Design assistants via chat.
  • Manually New: Directly fill in the complete configuration.

A practical initial team typically includes one Leader (planning, assignment, acceptance) and one or more execution members (coding, testing, review, documentation, etc.).

3. Build a Team

Click the floating Team button on the left side of the page, then click + in the workbench navigation bar:

  1. Add members from the assistant list (the same assistant can be added multiple times).
  2. Designate and only designate one Leader.
  3. Enter a team name and select a Workspace.
  4. Choose whether to allow users to communicate directly with regular members.
  5. Click Create and Start.

4. Describe the Goal to the Leader

Send the complete goal to the Leader. The Leader splits tasks, assigns members, receives progress, and accepts outputs. When team policy allows, you can also communicate directly with a regular member.

Collaboration Mechanism Briefly

The Leader creates tasks and assigns them to specific member instances via team tools; members execute tasks in their own Sessions, reporting in-progress, completed, or failed status; progress and results are automatically notified to the Leader. Members share a Workspace but do not share chat context. When an assistant joins a team, a configuration snapshot is generated; subsequent edits to the assistant template do not hot-update running members—old members must be removed and re-added to apply new configurations.

Use Cases and Considerations

Who is it for: Developers or teams within DSH who need to split models and tools by role, observe multi-channel Agent output in parallel, and collaborate on the same code Workspace. For example, architecture planning uses a high-capability model, coding uses a specialized model, commit messages use a lightweight model, while keeping the Leader’s context unfilled with implementation details.

Pre-use considerations:

  • The plugin runs with the current DSH process permissions. Before installing, review the GitHub source code and MIT license to ensure it meets your security and compliance requirements.
  • The “Changes” tab requires the selected Workspace to be a Git repository; regular directories can still be browsed for files.
  • Files selected from outside the Workspace are copied to .agent-team/uploads/ for stable reading by Agents.
  • MCP credentials are retained in the Harness Profile, and assistant templates only save the names of allowed Servers.
  • The SkillHub community directory (https://www.skillhub.cn) is an independent site with no official affiliation with DeepSeek / High-Flyer; the DSH ecosystem follows the “everything is a plugin” philosophy, and Agent Team is a workflow extension for multi-Agent collaboration within it.

Related Links

  • SkillHub Directory Page: https://www.skillhub.cn/plugins/limuyang2/agent-team
  • GitHub Repository: https://github.com/limuyang2/agent-team
  • npm Package: https://www.npmjs.com/package/@limuyang2/dsh-agent-team