DeepSeek harness minimal mode is the stripped-down agent preset: it switches off nearly everything the harness normally loads, leaving a very light, very fast agent built for quick, simple questions where you don't need files opened or tools run. It's one of four agent presets inside DeepSeek Harness, the free, MIT-licensed, open-source developer preview that shipped on 13 August 2026 and blew past 115,000 GitHub stars, making it one of the fastest-growing AI projects ever recorded.
📺 Watch: DeepSeek Harness: FREE 1 Hour Course!
🔥 Get the Agent OS as a free bonus: AI Profit Boardroom members get the full Agent OS zip, prompt libraries, daily tutorials and weekly live coaching calls. → Get inside
I recorded a full one-hour course on the harness in August 2026, and here's the pattern I keep seeing: almost everyone clicks past the preset picker on day one and never looks at it again. That's a mistake, because matching the preset to the job is free speed, and minimal is the preset most people don't even realise they have. If you haven't got the tool running yet, my DeepSeek Harness install guide covers setup; this page is about squeezing the most out of the lightest preset once you're in.
Where Minimal Fits Among the Four Harness Presets
Open the harness settings and you'll find an agent preset option with four presets. Each one is really a different amount of power handed to the agent, and you pick your preset when you start a new session.
- Standard is home base: balanced capabilities, good tool judgment, and built for the everyday work of summaries, drafts and research. It's where most of your time goes.
- Code unlocks heavier building powers for bigger multi-step technical jobs. I've broken that preset down properly in my DeepSeek code harness guide.
- Minimal strips the agent down to almost nothing: very light, very fast, and with far fewer moving parts.
- Creator builds brand-new plugins for the harness from a plain-English description of what you want.
The analogy I use in the course: running everything on a heavyweight preset is calling the whole team into a meeting, while minimal mode is just asking the person at the next desk. Most questions don't need the meeting. One housekeeping note while we're here: the preset picker is a separate dial from DeepSeek expert mode, which I've covered in its own guide, so don't muddle the two.
Want the full one-hour walkthrough? My complete DeepSeek Harness Masterclass — all four presets, the hidden config tricks and the exact workflows I use daily — sits inside AI Profit Boardroom, right next to my full Agent OS setup. Join us and skip the trial and error.
What DeepSeek Harness Minimal Mode Actually Strips Out
To see why minimal is so quick, you need one idea from under the bonnet: everything in this harness is a plugin. The sidebar is a plugin. The tools are plugins. The system prompt, the sandbox, even the agent loop itself — all plugins. A preset isn't some mysterious personality switch; it's a bundle of switched-on parts, and minimal is simply the bundle that switches most of them off.
That design buys you three real advantages:
- Speed. Fewer moving parts means the agent has less to load, less to weigh up and less to coordinate, so answers come back faster.
- Fewer tokens. The interface includes a context panel that breaks down exactly what's eating your tokens: system prompt, tools, messages. Strip the tools away and that overhead shrinks. The harness also caches, so re-reading existing context costs less than fresh material, and a lean session keeps the bill small to begin with.
- Fewer breakages. This is a v0.1 developer preview, so expect bumps. Plugins that work today can break next month, and the fewer parts you're running, the less there is to break mid-task.
The architecture underneath is called Cordis, and DeepSeek published an actual research paper on it: every action can be cleanly reversed, and pieces swap safely by design. That's the reason a preset can rewire the whole agent without anything falling over, and the reason this preset can strip so much out and still behave.
When I Use Minimal Mode (And When I Don't)
Honest guidance, exactly as I teach it in the course:
- Quick questions, small rewrites and sanity checks: minimal. "What does this error mean?", "Tighten this paragraph", "Is this subject line any good?" None of that needs an agent opening files or reaching for tools.
- Real file work: standard. The moment the job involves reading or editing your documents, you want home base and its tool judgment.
- Big builds: code. Multi-step technical jobs deserve the heavier preset.
- Making new tools: creator. Describe the plugin in plain English and let the harness build it for you.
I test this rather than assume it. On Goldie Bench, my own testing setup, one of my habits is running the same quick-question job across different configurations, and the lesson repeats: for simple asks, the stripped-down session gets you there with none of the ceremony. The whole point of DeepSeek harness minimal mode is that the person at the next desk answers faster than the meeting ever will.
Two related reads while you're mapping jobs to setups: my DeepSeek V4 Flash harness guide if speed is the obsession, and my Hermes agent DeepSeek breakdown for how the harness sits alongside the rest of my agent stack.
📺 Watch: DeepSeek Harness VS Claude Code
Customising Harness Presets, Including a Hidden Config Trick
The four boxes aren't locked. In settings you can view and edit how each preset is configured, and if your workflow doesn't fit any of them, you can draft a custom preset of your own. That's worth knowing before you contort your working day to match a default.
Then there's the deeper move I show in the course. On your machine the harness keeps a hidden folder called dsh. Inside it there's a folder called profiles, inside that one called web, and in there a file called cordis.patch.yml. Write any plugin's ID into that file with disabled true underneath it, save, reload the browser, and that piece of the harness is simply gone. It's minimal mode as a philosophy, applied by hand: you decide exactly which parts of the tool exist.
A word of caution, though. Because this is a developer preview, treat hand-edits as experiments. The clean-swap design of Cordis means pieces are built to be removed and restored safely, but I'd still change one thing at a time so you always know what did what.
📺 Watch: DeepSeek's New AI Harness Changes Everything
The Honest Limits of Minimal Mode
A stripped agent can't do file work, and minimal mode shouldn't be argued with about that. The parts that open files and run tools simply aren't loaded. If you catch yourself pasting half a document into the chat because the agent can't open the file itself, that's your cue: don't fight the preset, switch it. You choose a preset every time you start a new session, so the fix costs you seconds.
The other limit applies to the whole tool: it's early. A v0.1 preview that grew this fast will wobble in places, and minimal reduces your exposure to those wobbles without removing them entirely.
Quick Reference: Which Harness Preset for Which Job
| Job type | Preset |
|---|---|
| Quick question or sanity check | Minimal |
| Small rewrite or tidy-up | Minimal |
| Summaries, drafts, research | Standard |
| Anything touching your files | Standard |
| Big multi-step technical build | Code |
| Building a brand-new plugin | Creator |
DeepSeek Harness Minimal Mode: FAQ
Is minimal mode faster than standard?
Yes, and it isn't magic, it's arithmetic. Fewer plugins loaded means fewer moving parts and a smaller token overhead, which the context panel will show you in black and white. For simple questions you get snappier answers on jobs that never needed the extra machinery.
Can minimal mode open files or run tools?
No, and that's the entire point. Minimal strips the agent down to almost nothing, so it won't be opening files or reaching for tools. For that work, start a session on standard or code instead.
How do I switch between presets?
You pick the preset when starting a new session; the agent preset option in settings lists all four. There's no penalty for hopping between them, and I move between presets constantly depending on the job in front of me.
Is DeepSeek Harness free to use?
Yes. The harness shipped as a free, MIT-licensed, open-source developer preview, and I've answered the licensing question properly in is DeepSeek Harness free. The short version is that minimal, like the other three presets, is included.
Can I build my own preset?
You can. Every preset's configuration can be viewed and edited, you can draft a custom preset if the four defaults don't fit your workflow, and power users can go further still with the cordis.patch.yml file described above.
My Verdict on DeepSeek Harness Minimal Mode
Minimal mode is the most underrated of the four presets, and the reason is simple: people click past the preset picker on day one and never revisit it. But this harness is plugins all the way down, which means every session you run on a heavier preset than the job needs is paying for machinery it never uses. Match the preset to the job — quick stuff on minimal, real work on standard, big builds on code, new tools on creator — and you get speed for free.
Ready to go deeper than one preset? The full DeepSeek Harness Masterclass, the complete one-hour course this page is drawn from, lives inside AI Profit Boardroom, alongside my Agent OS setup and a community that pressure-tests this stuff every week. Join us and run the harness properly.
