DeepSeek Harness (dsh)

product
agent-harnessopen-sourceplugin-architecturedeepseek

DeepSeek Harness (dsh)

Type: Product (open-source agent harness / runtime)

DeepSeek Harness is an open-source agent harness published by DeepSeek at github.com/deepseek-ai/deepseek-harness, MIT-licensed, written in TypeScript, distributed as the npm package @deepseek-ai/dsh. The repository describes itself as "DeepSeek Harness: Everything is a Plugin" and states plainly that it "is currently in developer preview and is iterating rapidly." It shipped alongside the V4-Pro GA release on 2026-08-13.

Why it lands on this rung and not on models. It is not a model and carries no token price. It is the layer this rung exists to measure: the thing that decides how many tokens a job takes and whether it finishes. A model lab shipping the scaffolding around its own model, as a separate MIT artifact, is the harness layer being treated as a product rather than as an implementation detail — which is the rung's whole premise stated by someone with an incentive to state the opposite.

The architectural claim. Everything is a plugin — including the model adapter, the tool registry, the session log, the sandbox, and the agent loop itself. It is built on the Cordis plugin framework. If the agent loop is genuinely swappable, the harness stops being a fixed cost of using a model and becomes a variable a buyer can tune, which is the precondition for the cost-per-completed-task comparison this rung does not yet hold.

The model-adapter plugin is the strategically loaded one. A harness whose model adapter is a plugin is, by construction, not locked to DeepSeek's models. A lab publishing a model-agnostic harness under MIT is either building ecosystem gravity or conceding that the harness is where the value is accruing — and the two readings are not exclusive.

What is verified and what is not

  • Verified 2026-08-18, primary (the GitHub repository itself): MIT licence, TypeScript, @deepseek-ai/dsh, "developer preview", the plugin architecture claim, the Cordis basis.
  • Not verified: the exact preview version. Secondary coverage refers to "v0.1"; a specific tag (e.g. v0.1.0-rc.7) could not be confirmed from the repository and is therefore not recorded here.
  • Not verified, deliberately omitted: GitHub star counts. Independent secondary sources gave figures differing by ~50% within the same week, and the number is volatile enough to be worthless as a recorded fact.
  • No close-read file exists in kb/harnesses/raw/, and this entity is therefore not registered in _sources.json. Per the KB's own rule, a source with no close-read file has not been read; registering one would inflate the count. This page is a release record built from the repository's own front matter, not an ingest.

Open questions this raises for the rung

  • Does a swappable agent loop change measured cost-per-task, or only measured flexibility? This is the rung's biggest hole, and dsh is the first widely-available artifact that could be instrumented against it — same task, same model, loop swapped.
  • Does a model lab's own harness advantage its own models? A model-agnostic adapter is claimed; whether the default paths, prompts and retry policy are tuned to DeepSeek models is an empirical question, and a matched-model comparison would answer it.
  • Where does the harness sit when the model underneath it reprices? DeepSeek raised API prices 2–4.5× three days after publishing this (Aug 16). A free harness over a repriced model is the clearest available case of the harness layer and the token layer moving in opposite directions at the same provider.

Mentioned In

Related Entities

  • AlphaEvolve — the other lab-published agent system in this rung; AlphaEvolve is a closed service, dsh is MIT-licensed source
Related concepts