Preface

DeepSeek Harness (DSH) takes the “everything is a plugin” philosophy to its fullest: models, tools, and interface capabilities can all be assembled as needed. However, for practical development, a single-session Agent often falls short—code reviews require separate lines for security, performance, and product perspectives, competitor research demands parallel scraping, and documentation writing needs division between frontend and backend. Writing multi-Agent orchestration, task queues, and state machines from scratch is costly; integrating an independent Workflow engine adds deployment and maintenance overhead.

The community plugin dsh-agent-teams (npm package @nanmicoder/dsh-agent-teams) takes a path closer to DSH’s native experience: it allows the current session to act as a “captain,” automatically spawning resumable sub-Agent members, breaking down goals into dependent tasks, and providing persistent state, a shared scheduler, and a web activity panel. This plugin is maintained by NanmiCoder (author Relakkes) and is categorized as workflow in the SkillHub plugin repository. The GitHub repository NanmiCoder/dsh-agent-teams uses the MIT license; as of 2026-08-25, GitHub shows approximately 962 stars and 82 forks, with the latest npm version being 0.1.13. It should be noted that SkillHub is not affiliated with DeepSeek Harness officially—it is just one of the community-curated plugin directories.

What Is This

dsh-agent-teams is an AgentTeams collaboration plugin for DeepSeek Harness. In a nutshell: describe a goal in natural language, and the current DSH session becomes the captain, creating a team composed of sub-Agents, dividing work based on dependencies, and summarizing the final results.

The core pain point it addresses is: completing multi-Agent orchestration within Harness without needing an extra Workflow engine. The plugin comes with team protocols, 10 collaborative tools, file persistence, automatic task scheduling, and a real-time activity panel in the web GUI (including segmented progress, member tree, and an interactive task DAG).

Core Features and Highlights

Based on the official README and repository documentation, the verified main capabilities are as follows:

  1. Captain-style Delegation: The current session handles team creation, role assignment, and final result summarization, with a clear workflow.
  2. Resumable Members: Members are sub-Agents that can be awakened within DSH, suitable for multi-turn focused follow-ups rather than one-off scripts.
  3. Dependency-based Task Splitting: Tasks have clear states (e.g., running / idle / ready), and cannot be claimed if their preceding dependencies are incomplete, preventing out-of-order execution.
  4. Automatic Scheduling and Safe Handoff: Idle members automatically claim the next ready task; reassignment cancels the old attempt; after a cold process restart, the scheduler retries lingering open tasks.
  5. Direct Member Messaging: Members can communicate directly with teammates or the captain via a persistent mailbox, without the captain needing to relay each message.
  6. Real-time Web Activity Panel: The web UI displays team progress and the task DAG; team archives retain complete member and task history. Panel text switches between Chinese and English based on the Harness official locale.
  7. Slash Command Entry: Registers the /agent-teams command, selectable from the web GUI’s slash menu; in headless CLI, user messages starting with /agent-teams also trigger the protocol.
  8. Zero-interaction Team Creation by Default: Members default to using the captain’s current LLM routing; only when explicitly specifying heterogeneous division of labor (e.g., “backend uses model A, frontend uses model B”) is a specific provider + model assigned to a member.

Installation and Enablement

Before use, please confirm that DeepSeek Harness is installed on your machine. The plugin requires Node ^22.19.0 or >=24 (see repository package.json).

dsh plugin --profile web add @nanmicoder/dsh-agent-teams@latest

Method 2: Install from GitHub Source

The community directory deepseek-harness-plugin.com provides the command:

dsh plugin add github:NanmiCoder/dsh-agent-teams

For reproducible installation, pin a commit:

dsh plugin add github:NanmiCoder/dsh-agent-teams#<commit-hash>

Method 3: Local Source Build

git clone https://github.com/NanmiCoder/dsh-agent-teams.git
cd dsh-agent-teams
pnpm install
pnpm build
dsh plugin --profile web add .

After installation, check the combined configuration and start the web UI:

dsh --profile web --dump-config
dsh web

Security Note: DSH plugins run with the current dsh process permissions. Before installation, it is recommended to browse the GitHub source code and LICENSE to confirm they align with your trust policy.

Typical Usage Examples

Natural Language Team Formation

After installing and restarting DSH, you can describe goals directly in the conversation. For example, the official README sample:

Use AgentTeams to review commits after v0.5.3, dividing work from performance, security, and product perspectives, and output a summary report at the end.

The captain will create the team, add role members, split tasks, and coordinate execution, ultimately summarizing into a report.

Slash Commands

In the web GUI, select agent-teams from the slash menu or type directly:

/agent-teams research the pricing pages of three competitors

The command is retained in the chat history as a normal user message while triggering the AgentTeams protocol.

Optional Configuration

It works out of the box. If you need to override member behaviors in the profile, refer to:

- id: agent-teams
  config:
    stateDir: .agent-teams
    memberProvider: spawn
    memberModel: deepseek-v4
    memberMaxDepth: 1
    maxMembers: 8

Here, memberProvider refers to the sub-Agent runtime backend (spawn / fork), not the LLM provider. Setting slashCommand: false disables the deterministic activation of /agent-teams, leaving only natural language triggering.

Team state is saved by default in the .agent-teams/ directory under the workspace; the web panel reads this directory and merges it with real-time sub-Agent activity for display.

Applicable Scenarios and Considerations

Who Is It For

  • Developers already using the DeepSeek Harness web interface who want to keep “multi-role parallel + task dependencies + result summarization” within the same session.
  • Those needing code reviews, competitor research, documentation division, etc.—scenarios that can be split into multiple sub-tasks—but who do not want to build an orchestration system from scratch.

Usage Boundaries (As Explained in Official Documentation)

  • A captain can only lead one active team at a time.
  • State is file-persisted, operating serially within a single DSH process; simultaneous changes to the same team by multiple processes do not guarantee consistency.
  • The activity panel displays persisted state; occasionally, a model might complete work without updating the task status per protocol, requiring manual attention.
  • The plugin depends on peer packages like DSH sub-Agents and Web Client; please keep Harness and plugin versions relatively matched.

Conclusion

If you are already tackling complex tasks in DeepSeek Harness but still manually switching between multiple conversation windows to simulate “team division,” dsh-agent-teams is worth a try: one goal, one team, one activity panel, bringing multi-Agent collaboration into Harness’s native workflow.