DeepSeek Open-Sourced Its Agent Stack: What a Self-Hosted Agent Runtime Means for a Small Business
Synergised Consulting Ltd
Last Updated: 1 November 2026
DeepSeek Harness is an open-source agent runtime, an MIT-licensed program that sits between a language model and the tools it uses to actually do work. On 31 October DeepSeek also released a research prototype, the Super-Agent, built on top of it. Together they mean an owner can now self-host the entire runtime layer of an AI agent rather than rent one. The catch comes from the project itself: its own safety notice says the software has never had a security audit and should not be treated as production-ready.
What DeepSeek Actually Released
DeepSeek has open-sourced two related things. The first is Harness, shortened to dsh, a runtime layer that connects a language model to real work: reading files, running commands, holding session state, and calling tools. It is free to inspect, modify and self-host under the MIT licence. The second, released on 31 October, is the Super-Agent prototype, a separate repository showing how a proactive coding agent can be built on Harness. Both matter for the same reason: the plumbing that proprietary platforms charge for is now something a business can own outright.
According to The Next Web [1], DeepSeek describes the Super-Agent as a research prototype and "not an official DeepSeek product", and the public code differs from DeepSeek's own internal version. The Harness repository itself states plainly that it is a developer preview and that there will be compatibility-breaking changes [2]. This is a floor being laid, not a finished product being given away.
The Super-Agent Prototype: What It Does
The Super-Agent is a coding agent that works on one GitHub repository and can also listen to a team chat channel. When an approved person asks it for work, it opens a task, handles it in its own session, and replies in the original thread, sometimes with a pull request. A web panel lists the work that needs a human.
According to The Next Web [1], DeepSeek's own design reasoning is the interesting part: an agent acting for a whole team can waste the whole team's attention, so it acts only when asked, or when a team-defined rule allows it. That is a governance decision, made in public, by the team shipping the most aggressive open-source agent stack available. The design treats human approval points as a feature of the architecture rather than an obstacle to autonomy, which is the same conclusion any careful operator reaches after their first unplanned automation incident.
The design worth copying
The prototype's shape is worth naming, because it transfers to any agent a business builds, on any runtime. Requests arrive from a defined place. An authorised person, or a written rule, creates a work item. The item runs in its own isolated session and directory. Results return to the original conversation. A panel shows what is waiting for a human. Strip away the specific product and that is a five-part operating pattern: defined intake, authorised trigger, isolation, reported output, and a visible exception queue.
What It Costs, and What It Does Not
The runtime is free. That is the least important number in the decision.
Running an agent still costs the model tokens it consumes and the infrastructure it runs on. Self-hosting also transfers operational work: the updates, the security posture, the backups and the access control belong to whoever runs the server now. For context on DeepSeek's own scale and appetite, The Next Web reports the company has raised roughly $15bn ahead of a planned listing [1], so the giveaway of the tooling layer is a competitive move by a well-funded player, not an abandonware release. The runtime licence being MIT means no per-seat fee and no licensing surprise later, but it does not mean the total cost of running an agent well has changed much.
What Harness Is Not For
Harness is not for production workloads yet, by its own account. The project's SAFETY.md states that it is experimental developer-preview software that has not undergone a security audit, that it can execute model-generated code and load third-party plugins, and that its sandboxing and approval controls "do not guarantee isolation or prevent damage" [3]. It explicitly advises not relying on Harness as the sole security control for untrusted workloads, and recommends least privilege and a disposable environment instead.
The Super-Agent prototype is not a product either. DeepSeek says so directly, and warns that some ideas described in its design article are not implemented in the published code [1]. An owner reading a headline about an open-source super-agent should know that the honest version of the announcement includes the vendor telling you not to trust it yet.
It is also not an agent you install and use finished. As independent comparisons note, dsh is a harness you assemble from parts, while tools like Claude Code are finished agents you extend from the outside; Springbrand's comparison [4] frames this as the single distinction that decides most choices between them. That assembly work needs someone technical, which is a real cost even when the licence is free.
What This Means for a Founder in the First 100 Days
For a newly acquired business, the practical reading is not "switch to Harness this month". It is that the option to own agent tooling now exists, and that changes what to write down. A one-page tooling register, naming each agent tool the business uses, where it runs, what data it holds and what its licence allows, is worth producing now regardless of any decision. When a credible self-hosted alternative appears, the register is what turns "should we look at this?" into a quick comparison instead of a research project.
There is a second, quieter implication. In discovery conversations, we regularly meet owners who want AI capability and have the appetite to pay for it, but whose IT function is outsourced to a provider whose service scope simply does not include AI enablement, so there is no path to it through the existing relationship. An MIT-licensed, self-hostable runtime does not remove that blockage by itself, but it changes the conversation an owner can have: the question stops being "will you build us agents?" and becomes "will you host and secure this thing I can already inspect?"
The sensible posture for the next few months is parallel evaluation. Keep production work on tools that are finished and supported. Pick one low-stakes workflow, run it on a self-hosted harness on infrastructure you control, and learn what the operational load actually feels like while the stakes are small. If the runtime layer is genuinely becoming a commodity, the businesses that already know how to run it will adopt the mature version months before the ones still evaluating whether to look.
Sources
- [1] The Next Web: DeepSeek open-sources Super-Agent, a coding agent that reads team chats. DeepSeek Harness README: DeepSeek Harness: Everything is a Plugin. DeepSeek Harness SAFETY.md.
- [2] Background reading (qualitative guidance, no figures): Springbrand: DeepSeek Harness vs Claude Code comparison.
