The core idea of DeepSeek Harness (DSH) is “everything is a plugin”. When building complex workflows, multiple agents often need to collaborate in parallel. dsh-agent-teams-meta is a standalone plugin. It provides DSH with more robust Agent team management capabilities, supporting persistent teams, dynamic routing, task graphs, private mailboxes, token ledgers, and a Web sidebar panel.

What It Is

This plugin is maintained by kaka-in-home and is designed to address coordination challenges in multi-agent collaboration. It allows a lead agent to break down work among dedicated members. Each member has independent model routing, task context, and a private mailbox, and retains its working state within the same persistent session.

Core Features

  • Parallel members with real routing: You can assign different Providers and Models to members. Different members within the same team can use different models.
  • Persistent task coordination: Supports creating tasks, assigning owners, and defining dependencies. Members can only claim tasks for which they are eligible. Task status changes are validated and persisted.
  • Private persistent mailboxes: The lead agent and members exchange messages through mailboxes without interfering with each other. Each participant can only see the contents of their own mailbox.
  • Real-time model switching: You can switch a member’s model routing at runtime. The next request will use the new routing, while in-progress requests keep the original routing.
  • Useful failure reports: When a member request fails, the error information is passed to the lead agent as plain message text, so it does not disappear or become obscure.
  • Model catalog and routing validation: The plugin generates a catalog of the Provider/Models configured in the current DSH. When creating a member, routing validity is checked to prevent using unconfigured models.
  • Routing-based token ledger: Token usage records are based on the routing at the start of the request. Even if routing is switched later, the ledger still records usage under the initial routing.
  • Web visualization: Provides a read-only sidebar showing team progress, members, tasks, routing, and stages.
  • Toolset: Provides a teams_meta_* prefixed tool family, including create, add_member, create_task, claim_task, send_message, route_switch, status, close, and guide.

Installation and Activation

Install via DSH’s plugin command:

dsh plugin --profile web add dsh-agent-teams-meta

After installation, optional configuration can be done through cordis.patch.yml to customize the state directory or routing marker:

- insert:
    - id: agent-teams-meta
      name: 'dsh-agent-teams-meta'
      config:
        stateDir: .dsh-agent-teams-meta
        markerProvider: agent-teams-meta-route

Typical Usage

A typical collaboration workflow is as follows:

  1. Create a team: Define the team title and purpose.
  2. Add members: Specify member roles, responsibilities, Provider, and Model.
  3. Create tasks: Define concrete tasks and assign owners.
  4. Send instructions: Send messages through mailboxes to guide member actions.
teams_meta_create(title="Release audit", purpose="Review the release before publishing")
teams_meta_add_member(name="Reviewer", role="Find regressions", provider="deepseek-official", model="deepseek-v4-flash")
teams_meta_create_task(title="Review routing", owner="Reviewer")
teams_meta_send_message(to="Reviewer", body="Claim the routing task and report material findings.")

Runtime Requirements and Notes

  • Environment requirements: Requires Node.js ^22.19.0 or >=24. Requires the DeepSeek Harness Web Profile (dsh web) and at least one configured LLM Provider/Model.
  • Data storage: All data is stored in the .dsh-agent-teams-meta/ directory under the workspace. The plugin does not store prompts, model responses, credentials, or endpoint URLs.
  • Version note: The current version number is 0.1.1 (based on package.json).

Summary

dsh-agent-teams-meta provides DSH with advanced capabilities, progressing from simulated chat to real workflow collaboration. Through private mailboxes, task dependency management, and routing-based token tracking, it enables the lead agent to organize a team more effectively to complete complex tasks.