DeepSeek Harness Lets AI Generate and Remove Its Own Tools
DeepSeek Harness, an open-source environment presented by Two Minute Papers’ Károly Zsolnai-Fehér, is designed to let AI systems generate the interfaces, agents and workflows needed for a particular task rather than operate within a fixed workspace. Zsolnai-Fehér argues that its key mechanism is reversibility: each addition carries separate cleanup instructions, allowing task-specific components to be removed after use. A displayed Composio comparison shows the harness with a lower cost per successful task than Pi and Claude Code, though Pi has the higher reported success rate.
The point is to generate the tools for the task, not just run them
Károly Zsolnai-Fehér describes DeepSeek Harness as a free, open-source environment that gives an AI system “arms and legs”: a layer around the model through which it can do work rather than only return chat responses. Its defining promise, in his account, is not simply that users can configure an existing agent workspace. The workspace can generate the interface, modes, agents, and workflows a particular task requires.
That is the distinction Zsolnai-Fehér draws from existing harnesses such as Pi and OpenCode. DeepSeek Harness, he says, allows “every major part” to be rewritten. The user interface itself can change: its examples include a floating whale icon and a playable Snake game running beside the chat panel. The agents inside the harness can change as well.
A code-review mode, for example, is created from a request to inspect a codebase, find problems, and rank them by severity. In the shown result, the mode identifies duplicate accessible names across six HTML swatch links and marks the issue as important. The point is not the particular review finding; it is that the mode performing the review was made for that task rather than selected from a fixed menu.
It’s not you who rewrites the program, but the program rewrites itself.
The same pattern appears in the other demonstrations. A research mode checks statements in a draft against academic literature, marking the claim that more compute reliably improves model quality as supported while treating “Bigger models are always better” as contested. A local AI-lab dashboard monitors tokens per second, GPU utilization, and GPU-memory use. A video-production request produces a storyboard and shot plan for a 90-second explainer on diffusion models.
Zsolnai-Fehér’s broader claim is that an AI’s operating environment no longer needs to be fixed in advance. Instead, the system can construct the task-specific tools and interfaces through which it will carry out the work.
Self-extension depends on being able to remove what was added
Continually generating panels, agents, modes, and components creates an obvious maintenance problem: an environment that can alter itself also needs a way to avoid accumulating unneeded changes. Zsolnai-Fehér says the 88-page paper behind the harness addresses that problem by pairing each modification with the information needed to reverse it.
The accompanying notation and diagram pair an original action with a related cleanup function. Zsolnai-Fehér explains this as automatic retention of cleanup instructions: when the system adds a component, it also keeps what it needs to remove that component later. The resulting changes are intended to be reversible rather than permanent additions to the workspace.
With every change comes cleanup instructions, and the system remembers them automatically. New components can be safely removed. Everything is reversible.
His analogy is a coat check. Giving over the coat is the original action; the ticket is separate from the coat, but it carries the information needed to retrieve it. Likewise, the undo machinery can live alongside an action without modifying the action itself. The harness retains both the change and the means to undo it.
That separation matters to the case for a self-extending environment. A workflow-specific tool can be created when needed and removed when the work is finished, rather than becoming another irreversible part of the interface.
The displayed comparison favors lower cost, not the highest success rate
Zsolnai-Fehér also calls the harness lean and efficient, arguing that it can save time and money. A comparison credited on screen to “Data: @composio / Source: Composio agentic eval” makes a more specific case: DeepSeek Harness has the lowest displayed cost per successful task, but not the highest displayed success rate.
| Agent | Success rate | Cost per success |
|---|---|---|
| DeepSeek Harness | 66.7% | $0.028 |
| Pi | 70.0% | $0.031 |
| Claude Code | — | $0.074 |
Pi is shown with a 70.0% success rate, compared with 66.7% for DeepSeek Harness. But DeepSeek Harness is shown at $0.028 per success, below Pi’s $0.031 and Claude Code’s $0.074. The practical argument is therefore not that it leads every measure. It is that a system able to generate custom workflows may do so at a comparatively low displayed cost per successful task.
Zsolnai-Fehér says that only days after release, users had already created hundreds of plugins. The shown GitHub page for “Awesome DeepSeek Harness (DSH) Plugin” calls itself a curated list of plugins built on a framework “where everything is a plugin.” That description fits his larger framing: extensions are not peripheral add-ons to a fixed product, but the material through which the environment itself can be remade.