100 Hours Testing Deepseek Harness vs. Claude Code. What You Need to Know.
What this video is
⚡ a 18-minute video, readable in 60 secondsThe reviewer reaches no clean buy/wait/pass on DeepSeek Harness, an open source, free agent harness he tests side by side against Claude Code and Codex. He says it absolutely crushes those closed-source harnesses on customization and speed, but concludes it does not replace Claude Code for deep, complex work and that he trusts Claude Code's output more for a client or team deliverable. His mixed take: DeepSeek Harness is a promising but still-buggy developer preview, best used to offload cheap knowledge work when nearing usage limits elsewhere rather than as a full swap for Claude Code.
Verdict: mixed, reviewer says DeepSeek Harness does not replace Claude Code, especially for deep, complex work [10:52]
Key takeaways
- Verdict: reviewer trusts the Claude Code deliverable more and would be more willing to give it to a client or team than the DeepSeek Harness one [17:11]
- For: best suited to offloading knowledge work or research onto a cheap model in DeepSeek once nearing usage limits elsewhere [10:00]
- Price: the harness itself is completely free, you could run it 24/7 on a free model and never be charged, though inference is not free [05:53]
- Price: default billing is per-token API usage, you can't use a Claude subscription or Codex subscription directly [03:12]
+ 35 more takeaways
- Spec: DeepSeek Harness is an open source, free agent harness now in developer preview for agent harness developers worldwide, with source code included [01:56]
- Spec: DeepSeek's native models are DeepSeek R1, released last year, and DeepSeek V4 Flash and Pro, current [00:24]
- Spec: Opus, GPT, Kimi, or Meta models can also be added into DeepSeek Harness alongside its native models [00:59]
- Spec: DeepSeek V4 Flash and Pro are described as cheap and pretty solid, with Flash being cheaper [06:02]
- Spec: four modes are available before starting a prompt, Standard, PTC, Minimal, and Creator mode [03:57]
- Spec: Standard mode is a full coding agent with file editing, shell, file, and web search capabilities [04:07]
- Spec: Minimal mode is a two-tool coding agent, persistent bash and str_replace_editor, faster and cheaper for simple tasks [04:40]
- Spec: Creator mode is for building custom agent presets, plugins, and changing the UI [08:38]
- Spec: Settings then Plugins exposes a list of 140 plugins that can be toggled on or off [09:01]
- Spec: DeepSeek Harness has a Trajectory tab showing a granular step-by-step log of tool calls [12:51]
- Pro: open source design allows full customization, unlike closed-source harnesses like Claude Code [01:19]
- Pro: every capability, including tools, skills, sessions, sandboxes, storage, loops, and the UI, is a swappable plugin [02:02]
- Pro: no changes were needed to use Claude skills and context within this setup [05:45]
- Pro: much faster than Claude Code at finding things in his AI operating system [11:34]
- Pro: in a side by side test, DeepSeek took about one minute versus five-plus minutes for Claude Code on the same prompt [11:55]
- Pro: in a live test finding and summarizing a YouTube transcript, DeepSeek finished in about 50 seconds while Claude Code was still searching [12:26]
- Pro: on an Excel-report task, DeepSeek Harness finished in about three minutes versus about 17 minutes for Claude Code [14:12]
- Pro: customization is where DeepSeek Harness absolutely crushes, not even close to Codex or Claude Code [08:23]
- Pro: DeepSeek Harness's sugar-report output was more relatable and practical [16:42]
- Con: be careful with open source plugins since you don't know what might be inside them [02:18]
- Con: recommends having Claude Code or Codex review plugins before installing anything from the internet [02:24]
- Con: experienced a bug where the harness locked the mouse so it could only move within a small square area [07:40]
- Con: bugs with compression sometimes make it seem to forget what the speaker was talking about [07:51]
- Con: in long sessions, the harness has compaction bugs and context regressions [08:07]
- Con: Creator mode has a bug where clicking it instantly switches back to another mode [08:51]
- Con: DeepSeek V4 Flash and Pro can't see anything, no screenshot or vision capability, unlike Codex or Claude Code [06:09]
- Con: Claude Code came across as more conservative with its findings and facts, while DeepSeek Harness was sometimes too confident [16:57]
- Con: still feels like a developer preview [07:30]
- Vs: Claude Code, closed-source harnesses like Claude Code let you swap the engine but not rearrange internally, while DeepSeek Harness lets you swap seats, steering wheel, and change agent behavior under the hood [01:35]
- Vs: Claude Code, using Opus inside Claude Code via subscription gives better value than running Opus inside DeepSeek Harness [10:54]
- Vs: Claude Code, his custom skills were built for Claude Code, so switching harnesses changes how those skills are interpreted [17:24]
- Vs: Claude Code, on the sugar research task Claude Code's report was about 5,000 words versus DeepSeek's about 4,400 words [16:20]
- Vs: Claude Code, Claude Code's report cited 26 sources [16:27]
- Vs: DeepSeek, DeepSeek's report had at least 14 sources considered load-bearing [16:32]
- Vs: DeepSeek, verification note found 0 fabricated citations and 6 corrected or demoted citations flagged in place [16:34]
Shown on screen — grab and go
PROMPTThe AGENTS.md personal-identity system prompt
You're Nate Herk. Based on this workspace (Herk-2, your personal AI operating system): - YouTube creator focused on AI tools and automation. Your channel data and command center live in OtherWorlds/youtube-os/, and you have a Herk Highlights clips channel too. - Course creator running the AIS Masterclass (references/course-scripts/), with student resources and a community you reply to. - Author writing "Becoming AI Native" (OtherWorlds/Book/). - Business owner at Upp[...], finances tracked in [...]
merged from repeated OCR reads of the AGENTS.md context-injection panel shown 05:02-05:12; first four bullets are legible, later bullets were too obscured to reconstruct and are omitted. shown at 5:04
How this brief was shaped: Evaluation (review / comparison / unboxing) · confidence Medium
Creator states his honest opinion after a week of use, compares it directly to Claude Code, and OCR shows the DeepSeek Harness GitHub repo confirming the product being reviewed. Transcript also breaks down specific modes like PTC and minimal mode, which reads as review depth rather than a pure side by side buyer comparison.
The lens sets this brief's structure, never its facts — every claim is held to the same citation and fact-check standard.
Jump to a moment
Their links, sorted & clickable
🏛️ Communities & courses1
My FREE resources🤝 Work with them1
My playbook for growing a $1M AI agency🛠️ Tools they use1
FREE MONTH voice to text📢 Sponsored / affiliate1
Code NATEHERK for 10% off VPS (annual plan)💼 Sponsorship & business1
Sponsorship InquiriesWas this brief useful?