DSHGuide

Contents / Chapter 1

Chapter 1 — Meet DSH: Models, Harnesses, and Agents

4 min read · 867 words · 11 sections

What You Will Learn

  • Explain DSH in one sentence;
  • Tell the difference between a model, a harness, and an agent;
  • Understand Session, Workspace, Tool, and Skill;
  • Choose between Desktop, Web, and Headless.

1.1 A Useful Formula: Agent = Model + Harness

Start with one formula:

Agent = Model + Harness

Term Think of it as It handles
Model The brain Understanding, reasoning, writing, coding
Harness Hands, eyes, workbench, rulebook Files, commands, web access, sessions, permissions
Agent The assistant at work Model + Harness + current context

A model alone is like a brilliant person locked in a room. It can think and talk, but it cannot see your files or run your programs.

The harness opens the room. It gives the model files, a terminal, a browser, tools, memory, and rules. That is when chatting becomes doing.

DeepSeek Harness, or DSH, is that harness.[1]

1.2 What DSH Is — and Is Not

It is

  • A workbench for running Agents;
  • A host that composes models, tools, skills, and plugins;
  • A platform with Desktop, Web, and Headless entry points;
  • An extensible system where almost everything is a plugin.

It is not

  • A model. DSH does not contain the model's brain; you connect a model service.
  • Just a chat app. Its focus is reading, writing, executing, and delivering.
  • Only for programmers. Everyday tasks can be done in natural language.
  • Safe by default. It can take real actions; permissions and review matter.

In one sentence:

DSH is the Agent's workbench, not its brain.

1.3 Five Words You Will See Everywhere

Agent

An Agent is not just a model. It combines the model, the current workspace, the current session, available tools, and the current mode.

The same model can behave differently in a different workspace or mode.

Session

A Session is one continuous conversation.

  • Within a Session, the Agent remembers what was said;
  • A new Session does not automatically inherit every previous message;
  • One workspace can have many Sessions.

Think of a Session as the minutes of one meeting.

Workspace

A Workspace is a folder. The Agent reads, writes, and runs commands inside it by default.

  • It is the project boundary;
  • Many permissions are scoped to it;
  • Do not start by pointing it at your entire home directory.

Tool

A Tool is an action the Agent can take:

  • Read a file;
  • Write a file;
  • Edit a file;
  • Run a terminal command;
  • Search the web;
  • Start a subagent;
  • Present a deliverable.

Tools are the Agent's hands.

Skill

A Skill is an instruction manual. It tells the Agent:

  • When to use a certain approach;
  • What steps to follow;
  • What pitfalls to avoid;
  • Which references or scripts are available.

A Skill is knowledge. It does not execute actions itself; it helps the Agent use tools reliably.

Model is the brain, Tool is the hand, Skill is the procedure, Harness is the workbench.

You will meet Preset, Plugin, Bundle, and Profile later. Chapters 7 and 8 cover them.

1.4 Why a Harness Is Necessary

A model is smart, but real work needs three things a model does not have by itself:

  1. Perception: seeing files, folders, web pages, and errors;
  2. Action: writing files, running commands, calling tools;
  3. Constraints: knowing what is allowed and how to verify the result.

Suppose you say:

"Combine the three articles in notes/ into one summary."

The model can write a summary, but it needs the harness to:

  • Find notes/;
  • Read the three files;
  • Write summary.md;
  • Verify that the file exists;
  • Stay inside the permissions you granted.

Without a harness, none of those actions happen.

1.5 Three Entry Points: Desktop, Web, Headless

Entry point Best for What it feels like
Desktop Almost everyone A graphical app; easiest to start
Web Browser access, remote servers A web page with a similar interface
Headless Automation, scripts, batch jobs A one-shot command-line run, no GUI

All three share the same core ideas: Agent, Session, Workspace, Tool, Skill, Preset, and Plugin.[2]

This book is Desktop-first. Web and Headless differences appear in small sidebars.

1.6 How to Read This Book

  • Everyday user: Chapters 1–7 cover most needs.
  • Installing plugins and skills: add Chapter 8.
  • Building plugins: add Chapter 9.
  • Troubleshooting: go straight to Chapter 10.

Do not try to memorize every term. Use first; understand later.

Try It

Write down three things you would like an Agent to do. For each, note:

  1. Which files it needs to read;
  2. Which files it needs to write;
  3. Which commands it needs to run;
  4. Which actions should ask you first.

If you cannot answer yet, Chapter 3 will help.

Common Pitfalls

  • Treating DSH as a model: it is a workbench; connect a model separately.
  • Expecting the Agent to remember everything: memory outside a Session must be designed.
  • Giving it your whole disk on day one: start with a small folder.

Summary

  • Agent = Model + Harness.
  • DSH is the harness, not the model.
  • Session is a conversation; Workspace is a folder; Tool is an action; Skill is knowledge.
  • Desktop is the easiest entry point; Web and Headless are different surfaces, not different products.

References

  1. DeepSeek Harness official site
  2. DeepSeek Harness GitHub repository

All links are public sources; product behavior may change between versions.

↑