DeepSeek has added a Creator Mode to DeepSeek Harness, the app it publishes on GitHub, the company’s Harness account said on X on Monday, 2 November. The post says users can “Turn ideas into plugins that you can use, from a design workspace to a sandbox game or a 3D LLM visualizer”.
The project’s GitHub repository showed 269,397 stars and 32,719 forks when we read it at 17:37 UTC, and it was created on 13 August. By our arithmetic, that is about 3,300 stars a day over its roughly 81 days.
The key insight: DeepSeek is pitching Harness as an app users shape themselves, with Creator Mode for building their own plugins, while its own design article argues the limit on running many agents is human attention, not the model.
The post opens with “Your Harness. You Define It.” and ends with “What will you create?”. It does not give a release version, pricing or availability details. The account’s bio reads “Everything is a plugin”, and the repository describes the app as “DeepSeek Harness: Everything is a Plugin”.
The post had 525,396 views and 10,429 likes when we read it at about 17:30 UTC, against 60,421 followers for the account. Bloomberg also ran the launch under the headline “DeepSeek Boosts AI Agent Offering With New Creator Mode”; we have not read the Bloomberg article.
The latest release listed on GitHub is dsh-v0.2.1-rc.5, published on 30 October, and the repository is under the MIT License.
Task completion time at p95 in DeepSeek’s Super-agent experiment (15.7 minutes for the Router-plus-Workers design vs 29.5 minutes for a single session with a neutral prompt), as reported in the DeepSeek Harness X Article of 1 November 2026, which calls the results indicative only; plus the Harness repository’s 269,397 GitHub stars at 17:37 UTC on 2 November and about 3,300 a day since 13 August (our arithmetic).
The Design Article Behind It
A day earlier, on 1 November, the same account posted a long X Article titled “Exploring Super-agent Design: Towards a More Autonomous Agent System”. It says the original is in Chinese and the English version is a machine translation, and that it “is not an introduction to existing features of DeepSeek Harness”.
The article describes a two-level design. A long-lived Router layer handles all human interaction and coordination, while context-isolated Workers run tasks. It says “a person can usually follow only 3 to 5 active sessions at once”, and that with current designs “multi-tasking is limited by human attention”.
It lists four traits it wants in a Super-agent: continuous, productive, proactive and persistent. A separate Watchdog session, it says, handles infrastructure failures and judges whether each Worker is making progress.
What the Experiments Showed
The article says its tests used “synthetic workloads and simplified environments” and that the results are “indicative only”. Each of four designs ran 5 times on about 300 chat and GitHub CI messages, about 20 of them work requests, with coding simulated by a tool that simply consumed time.
By the article’s figures, the Super-agent design completed tasks in 12.4 minutes at p50 and 15.7 minutes at p95. A single session with a subagent tool and a neutral prompt did not call a subagent in any of the 5 runs, and its p95 reached 29.5 minutes. With a delegation-first prompt, the model delegated in 4 of the 5 runs, the article says.
Turning off pass-through, which routes follow-up events straight to the subscribed Worker, left completion times about the same, according to the article. But the Router made about 30% more decisions per run, and cache-read tokens rose from 23.2M to 31.8M, “about 37% more”.
In a separate stress test, the article says, raising Trigger arrivals from about 1 to about 40 per minute lifted the average batch size from 1.0 to 2.8. Time per request rose from 2.5 s to 3.9 s, while the time per decision fell from 6.4 s to 1.6 s.
As we read it, the two posts make one argument: Creator Mode moves plugin building to the user, and the Super-agent article moves session management from the user to the system.
As we read the experiment, the cost is tokens rather than time: the article says the Super-agent design uses more cache-read tokens per run than single-session designs, trading tokens for shorter and more stable completion times.
DeepSeek Harness X Article, 1 Nov 2026 (machine translation)
“With these designs, a single task is limited by model capability, but multi-tasking is limited by human attention”
Three Implications
FOR DEVELOPERS DeepSeek’s post says Creator Mode turns ideas into plugins, with examples from a design workspace to a sandbox game or a 3D LLM visualizer.
FOR AGENT BUILDERS The article says a single session with a neutral prompt did not delegate in any of 5 runs, while the two-level design kept p95 at 15.7 minutes in its synthetic tests.
FOR TOKEN BUDGETS The article says turning off pass-through raised cache-read tokens from 23.2M to 31.8M, about 37% more, with similar completion times.
The framework reads: “The surface, not the model, governs autonomy.”
As we read it, DeepSeek’s article makes a similar case from the builder’s side, crediting the system design rather than the prompt for when work gets delegated. The framework was published in July 2026 and does not discuss DeepSeek Harness.
Why AI systems earn autonomy in steps, and where human approval stays in the loop.
What Is Not Established
DeepSeek’s post does not say whether Creator Mode is in the current GitHub release, which models it uses or what it costs. The Super-agent article describes architectural ideas, not shipped features, and its experiment figures are DeepSeek’s own on synthetic workloads; they have not been independently verified. GitHub stars measure interest, not usage. We did not contact DeepSeek.
This Week In The AI Supercycle
Anthropic’s Reality Check, The Capital Squeeze, & The New World Order
This week’s issue: for two years the AI build-out was priced as a growthstory. This week it started being priced as a credit story.
DeepSeek says Creator Mode lets Harness users turn ideas into their own plugins, and the app’s repository had about 269,000 GitHub stars on 2 November, roughly 3,300 a day since its 13 August creation by our arithmetic.
A note on sourcing. The Creator Mode details and quotes are from DeepSeek Harness’s 2 November post on X; the Super-agent design and experiment figures are from the account’s 1 November X Article, a machine translation of a Chinese original whose experiment results DeepSeek calls indicative only. Repository figures are from GitHub at 17:37 UTC. The stars-per-day figure is our arithmetic, and we did not contact DeepSeek.
Gennaro Cuofano is a CRO and tech executive who has worked in AI since late 2015, starting with bringing NLP, natural-language generation, voice and chatbot products to the marketing industry. His background is in law and finance: he holds a Master's degree in Law and an International MBA with an emphasis on corporate finance (LUISS Business School and the University of San Diego, 2012), and worked as a financial analyst at a real-estate investment firm in San Diego and as an assistant controller. His work focuses on business model strategy, business engineering and, more broadly, structural analysis: how companies make money, read from their own filings. He created FourWeekMBA and leads research there and at The Business Engineer, his newsletter on AI and business strategy, with over 95,000 subscribers and more than 1,000 published analyses. His writing has also appeared on Entrepreneur (for example, How to Digitize a Family Business in Europe) and HackerNoon. Find him on LinkedIn and Substack. See how we source.
