Preface

The design philosophy of DeepSeek Harness (DSH) is “everything is a plugin”. When handling complex tasks, multiple agents often need to specialize and collaborate, rather than a single session handling all the logic. The dsh-agent-teams plugin solves this pain point. It transforms the current DSH session into a “captain”, responsible for breaking down goals, assigning roles, and coordinating a team composed of persistent members.

Plugin Overview

dsh-agent-teams is an open-source plugin (MIT License) maintained by Wuxie233. It allows users to use natural language instructions to have a captain agent organize a multi-agent team. The plugin provides team protocols, ten coordination tools, persistent state, and a real-time web UI, and can run without an additional standalone workflow engine.

Core Capabilities

  1. Captain-led delegation: The current session automatically becomes the team captain, responsible for creating the team, assigning roles, and consolidating the final results.
  2. Persistent members: Members are resumable DSH subagents that can be awakened for focused follow-up turns.
  3. Dependency-aware tasks: Tasks use explicit state transitions and cannot be claimed until their dependencies are completed.
  4. Direct messages: Members can send persistent email-like messages to teammates or the captain without a relay.
  5. Real-time activity panel: The web UI displays roles, current work, unread messages, task dependencies, and archived history.

Installation and Enablement

Before using it, make sure DeepSeek Harness is installed. The installation command is as follows:

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

After installation, verify the configuration and restart the web service:

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

Typical Usage

Enter natural language instructions directly in a DSH session, for example:

Use AgentTeams to review the commits after v0.5.3 from performance, security, and product perspectives. Return one consolidated report.

The plugin will automatically decompose the task, assign members with different perspectives (performance, security, product), and coordinate the generation of a consolidated report.

Notes and Boundaries

  1. Single-captain limitation: Only one captain can lead one active team at a time.
  2. Wake-up mechanism: Members act only after being explicitly awakened; emails remain persistent while participants are idle.
  3. Concurrency limitation: State is file-based and is serialized within each DSH process. Multiple processes editing the same team simultaneously are not coordinated.
  4. State synchronization: The activity panel reports persistent state as-is. Sometimes the model may complete a task without updating the expected task state.

Conclusion

dsh-agent-teams provides a lightweight approach to building multi-agent systems inside DSH. It simplifies complex multi-agent collaboration into natural language interaction, making it suitable for scenarios that require structured task decomposition and team collaboration.

For more information and source code, see: GitHub repository.