Best DeepSeek Harness Workflow and Automation Plugins (2026)

DeepSeek Harness (dsh) ships an official workflow engine (ctx.workflowEngine) and a "goal" subsystem for persistent, multi-round objectives, but multi-agent team orchestration, scheduled cron-style runs, and domain-specific research loops are all community plugin territory. Here are eight workflow and automation plugins from FindHarness's curated index, ranked by GitHub stars as of August 2026.

The picks at a glance

Plugin	Stars	What it does	Link
dsh-agent-teams	253	Multi-agent team orchestration	Plugin
dsh_workflow	54	Generatable, governable, resumable workflow layer	Plugin
mstar-harness	43	Skill-driven harness/loop engineering workflow agent	Plugin
dsh-automation	32	Scheduled coding runs with auditable history	Plugin
dsh-suite (plugin-team-board)	16	Shared multi-agent task board	Plugin
dsh-deep-research	10	Deep-research orchestrator on the official workflow engine	Plugin
dsh-science	9	ReAct research loop + 10 genomics/bioinformatics skills	Plugin
dsh-sentinel	6	Condition-driven wakeup: file/command/http/webhook watches	Plugin

All eight come from the workflow-automation category in FindHarness's curated (source: awesome) index. None have been security-reviewed by FindHarness — read the plugin security checklist before installing, especially before granting any of these broad multi-agent or scheduling permissions.

What's official vs. what's a plugin here

Two pieces of this category are genuinely first-party dsh concepts, worth separating from the plugins below:

• Workflow engine (workflow/workflow, workflow/tool-ralph): a general-purpose workflow engine exposed as ctx.workflowEngine, plus a "Ralph loop" tool — a pattern where each round runs as a fresh, no-history child session, with state carried forward only through a shared workspace and a structured handoff report (status/summary/evidence/next steps/blocker).
• Goal (goal/goal, goal/command-goal): a persistent objective attached to an existing session, with active/paused/blocked/complete states and a /goal human command.

Subagents are also official — dsh's ctx.subagents registry supports delegating to in-process children, Claude Code, Codex, or an external ACP process. dsh-agent-teams (below) builds a specific multi-agent team pattern on top of that underlying capability rather than replacing it.

The eight plugins

1. dsh-agent-teams — multi-agent team orchestration

github.com/nanmicoder/dsh-agent-teams · 253 stars · MIT · npm: @nanmicoder/dsh-agent-teams

By a wide margin the most-starred workflow plugin. "AgentTeams multi-agent teams" is the entire description in FindHarness's index — for the specifics of how teams are defined and coordinated, check the repository's own README, but the star count alone signals this is the community's default answer to "how do I run more than one agent together" beyond dsh's raw subagent delegation primitives.

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

2. dsh_workflow — a governable workflow layer

github.com/icetomoyo/dsh_workflow · 54 stars · MIT

Described as "UltraCode-style multi-agent orchestration": a workflow layer that's generatable, savable, governable, observable, and resumable. The emphasis on governance and observability distinguishes it from a simple task runner — it's aimed at workflows you want to inspect and control mid-run, not just fire and forget.

dsh plugin --profile web add github:icetomoyo/dsh_workflow

3. mstar-harness — skill-driven loop engineering

github.com/btspoony/mstar-harness · 43 stars · MIT · npm: @mstar-harness/dsh

A "skill-driven harness/loop engineering workflow agent plugin." The framing around loop engineering suggests overlap with dsh's own Ralph loop concept, but as a plugin built specifically around skills — check the README for how it composes with or differs from the official workflow/tool-ralph package before assuming equivalence.

dsh plugin --profile web add @mstar-harness/dsh

4. dsh-automation — scheduled coding runs

github.com/titanwings/dsh-automation · 32 stars · MIT

Scheduled coding runs in fresh agent sessions with auditable history. This is the closest thing on this list to a "cron for dsh" — each scheduled run starts a new session rather than reusing an existing one, and the history is kept auditable, which matters if you're running unattended tasks and want a record of what each run actually did.

dsh plugin --profile web add github:titanwings/dsh-automation

5. dsh-suite (plugin-team-board) — a shared task board

github.com/whyihaveyou/dsh-suite (packages/plugins/plugin-team-board) · 16 stars · MIT

A shared multi-agent task board — create, claim, transition, and query tasks — implemented over a Cordis service key so multiple agents/plugins can coordinate through it. This solves a narrower problem than dsh-agent-teams: it's a coordination primitive (a shared board) rather than a full team-orchestration framework, and could plausibly be used underneath a custom multi-agent setup. Note this plugin lives inside a monorepo at packages/plugins/plugin-team-board — check the repository's own README for the exact install path.

dsh plugin --profile web add github:whyihaveyou/dsh-suite

6. dsh-deep-research — research orchestration on the official engine

github.com/omdsh-dev/dsh-deep-research · 10 stars · MIT

An adaptive deep-research orchestrator explicitly built on the official workflow engine rather than reimplementing its own loop logic. Worth calling out for exactly that reason: if you want a working example of what a plugin looks like when it's built on top of ctx.workflowEngine rather than around it, this is one of the few in the curated index that says so directly.

dsh plugin --profile web add github:omdsh-dev/dsh-deep-research

7. dsh-science — a domain-specific research workbench

github.com/biociao/dsh-science · 9 stars · MIT

A "Claude Science-style research workbench": a ReAct research-loop engine exposing research_* tools, versioned artifacts with provenance tracking via artifact_* tools, and ten packaged science skills covering genomics, pathogens, and bioinformatics. This is the most domain-specific plugin in this list — most useful if your work actually touches those fields, but a good reference for what a fully-fleshed research workflow plugin (loop + artifacts + skills) looks like even if you don't.

dsh plugin --profile web add github:biociao/dsh-science

8. dsh-sentinel — condition-driven wakeup

github.com/fuhefei/dsh-sentinel · 6 stars · BSD-3-Clause

Durable watches on files, commands, HTTP endpoints, processes, or webhooks that wake the agent when a condition fires. This is a different automation shape from dsh-automation's scheduled runs — instead of "run every N hours," it's "run when X happens," which matters for tasks that are genuinely event-driven (a file changing, a webhook firing) rather than time-driven.

dsh plugin --profile web add github:fuhefei/dsh-sentinel

Choosing between the multi-agent options

Three plugins here touch multi-agent coordination in different ways:

• dsh-agent-teams — the most-adopted, general "run a team of agents" plugin.
• dsh_workflow — a governable workflow layer, more about structure and observability than team composition per se.
• dsh-suite (plugin-team-board) — a narrower coordination primitive (a shared task board) rather than a full framework — useful if you're building something custom and just need the board.

Which one fits depends on whether you want a ready-made team pattern (dsh-agent-teams), a governed workflow structure (dsh_workflow), or a coordination building block to compose yourself (plugin-team-board).

FAQ

Do I need any of these plugins to use subagents in dsh?

No — subagent delegation (ctx.subagents, including delegating to Claude Code or Codex) is an official dsh capability, not something these plugins add. What the plugins here add is orchestration on top of that primitive: team patterns, scheduling, and governance.

What's the difference between the Ralph loop and these workflow plugins?

The Ralph loop is dsh's own official pattern for a specific style of iteration: each round is a fresh child session with no inherited history, coordinating only through a shared workspace and a structured handoff report. The plugins here (dsh_workflow, mstar-harness) may implement similar or different loop patterns — check each README rather than assuming they follow the Ralph shape.

Is dsh-automation the same as a cron job?

Functionally similar — scheduled coding runs in fresh sessions — but it's a dsh plugin, not the OS-level cron. It integrates with dsh's session and history model in a way a raw cron job calling dsh --profile headless wouldn't automatically get (see DeepSeek Harness Headless Mode for the CLI-only alternative).

Which of these plugins are npm-published vs. GitHub-only?

dsh-agent-teams (@nanmicoder/dsh-agent-teams) and mstar-harness (@mstar-harness/dsh) are published to npm — a prebuilt install, no allowBuilds prompt. The other six are GitHub-source installs, which may trigger the allowBuilds prompt on first install if they need a build step.

Can I combine dsh-deep-research or dsh-science with the goal subsystem?

Nothing in the descriptions rules it out, and dsh-deep-research is explicitly built on the official workflow engine that goals also interact with — but this isn't documented behavior we can confirm, so test it in a throwaway profile before relying on the combination.

Next steps

For the official concepts referenced throughout this article — subagents, the workflow engine, and Ralph loops — see DeepSeek Harness Subagents. Browse the full category at /categories/workflow-automation, or see 10 Best DeepSeek Harness Plugins in 2026 for our broader picks.