🧠 Model + Harness = Agent
The one idea that makes the whole course click: "An AI model is a brain — it can think, but it can't touch anything. It can't open files, remember last Tuesday, use a browser, or run a program. A model is a very smart voice in the jar. The harness is everything wrapped around that brain — the hands, memory, workspace, and rules." 2:04
The launch facts: August 13th developer preview, MIT license, 115,000+ GitHub stars (fastest-growing AI project ever), 300+ community plugins in days, led by Tianyi Cui (joined DeepSeek March 2026 from Jane Street), 19-person team from a May proof-of-concept to August release. 0:56
⚡ Level 1: Install & First Task
The install is intentionally friendly — and web-first, unlike every other terminal-only harness: 4:00
| Step | Action |
|---|---|
| 1 | Install Node.js (nodejs.org) |
| 2 | Open terminal (Mac: cmd+space "terminal"; Win: PowerShell) |
| 3 | One install command → runs locally at localhost:3080, opens in your browser |
| 4 | Connect a brain — DeepSeek API key (platform.deepseek.com), or OpenRouter/OpenCode for free models |
| 5 | Add a workspace — one clean folder (the "room" the agent works in) |
| 6 | Write your first real task in plain English |
The workspace is a safety + focus boundary: "It can read, create, and change files in that folder — nothing outside it. Don't point it at your whole desktop; make one clean folder. It keeps things safe and keeps the agent focused — digging through thousands of irrelevant files costs tokens and time." 7:40
His advice: "Give it real jobs from day one — don't vibe-code a snake game. Look at what you spend time on daily and automate that. One automation per week." 12:20
🔍 Trajectory View
"DeepSeek made every step visible." The trajectory view lays out every single step the agent took, graphically clickable and in order: the initial system prompt, your message, the context loaded, its actual thinking (reasoning in full), every tool call, every command run, the exact payloads and results, and how long each step took. 13:51
This is "the complete opposite of closed labs, where the model's thinking is hidden from you." You can also export any session as a zip — a complete, permanent log in standard format, "so you can check the work without redoing it. Review the trajectory the way a manager reviews a report." 17:00
🧩 Level 2: Everything Is a Plugin
DeepSeek's GitHub headline is three words: everything is a plugin — and they mean it literally. "The sidebar is a plugin. The tools are plugins. The system prompt is a plugin. The sandbox is a plugin. Even the agent loop itself — the core think-act-check-repeat cycle — is a plugin. Every organ of this machine can be switched off, swapped out, or replaced without breaking the rest." 20:26
Under the hood is Cordis — and DeepSeek did what labs almost never do: published an actual research paper on the design. The core idea, simply: "Every action can be cleanly reversed. Add a piece, and there's a guaranteed clean way to remove it — no leftover mess, no side effects. That's the reason 'everything is a plugin' actually works. The swapping is safe by design, mathematically — not held together with tape and hope." (Also "very Jane Street: build the boring, rigorous foundations first so the exciting stuff on top doesn't fall over.") 23:18
The four agent presets: 23:43
| Mode | Purpose |
|---|---|
| Standard | Home base — balanced, good tool judgment, everyday summaries/drafts/research |
| Code | Heavier building powers for complex multi-step technical jobs |
| Minimal | Strips down to almost nothing — light, fast, no tools (quick questions) |
| Creator | Build new plugins by prompting |
🛠️ Creator Mode
"Creator mode exists for exactly one job: building new plugins by prompting. Describe a tool you wish existed in one plain English sentence, and the harness builds it live inside your session and switches it on." 25:10
Plugins live in-session by default, but can be persisted to become permanent parts of your harness. You can stop, start, edit, and delete them anytime. 28:10
🗂️ The Hidden Config File
"The buried setting almost nobody will find": on your machine there's a .dsh folder → profiles/web → cordis.patch.yml. Open it in any text editor. 30:30
disabled: true, save, reload, and that piece is gone. You can edit the harness and mold it however you want — or get the agent to do it for you. That's what they mean by editing the harness." 30:35
Plus: community plugin marketplaces (e.g. deepseekcode.com/plugins) are already appearing, and there's a plugin search/install flow in settings. "It creates almost an app store for DeepSeek — everything is a plugin, every piece can be swapped in and out." (With the standard caution: vet anything you install from GitHub.) 32:10
🔗 MCP, ACP & agents.md
"One line in the docs matters more than the whole launch announcement": DeepSeek Harness supports MCP, the Agent Client Protocol, and reads agents.md / claude.md files. 33:08
| Standard | What it does |
|---|---|
| MCP | A standard plug socket for tools — any agent connects to any tool without custom wiring |
| Agent Client Protocol | A standard way for agent software to talk to the apps around it |
| agents.md / claude.md | Instruction files — rules, writing voice, projects, where things live, what to never do |
🏗️ Level 3: The Agentic OS
The reframe that dissolves tool fatigue: "The picture in your head might be a treadmill — every new tool is another thing you must master, so more tools means falling behind. But mastery of individual tools was never the skill. The skill is running one system that new tools plug into. When you have the system, a new tool isn't homework — it's a free upgrade delivered to your door." 36:30
His "Mission Control" dashboard sits every agent side by side — SEO workflow, video agent, loop system, kanban, memory, and DeepSeek Harness — all in one place. But the thing that makes it a system rather than parallel apps is shared memory via Obsidian: 38:20
The vault structure is deliberately simple: Daily / Projects / Research / Outputs, with an agents.md at the top. "Complicated vaults don't work. A handful of folders: daily notes, one note per active project with status, research, and output drafts. Every agent that opens the vault gets briefed automatically. If you can organize a filing cabinet, you can build this." 39:10
The payoff: DeepSeek Harness "knows my brand from day one because it's pulling context from Obsidian — what we work on, who we work with, our FAQs." Multiple agents can code into the same folder simultaneously (six conversations building features at once), and the memory updates as agents work — "a positive feedback loop." 42:20
✅ Key Takeaways
- The harness, not the model, is the work. A model is a voice in the jar; the harness is the body — what it can see, do, and how it recovers.
- Everything is a plugin — literally. Even the agent loop and the system prompt. The Cordis runtime makes every change reversible, so swapping parts is mathematically safe.
- Creator mode is the killer feature. Describe a tool in one sentence, approve it, and it appears. "The interface becomes clay."
- Trajectory view is the trust mechanism. Every step, visible and clickable — you can diagnose exactly where a job went wrong.
- Compatibility is a strategy. MCP + ACP + agents.md/claude.md means DeepSeek speaks the language every agent already speaks.
- Level three is the real endgame. The skill isn't mastering tools — it's running one system they plug into, with Obsidian as shared plain-text memory.
- It's version 0.1. Powerful, but a developer preview — build it into a system, not as a finished dependency.
🔗 Resources & Links
- 📺 Original video — Julian Goldie's full course
- 🔌 DeepSeek Harness — the repo