What is DeepSeek Harness?

DeepSeek Harness (DSH) is a Cordis-based agent runtime where every capability — models, tools, sandboxes, UI — is a swappable plugin. It lets you compose AI agents from a rich ecosystem of community-built components.

Oh! dsh is an independent expert guide for DeepSeek Harness. We provide curated plugin reviews, real-world scenario guides, and bilingual documentation to help you make the right choices for your DSH projects.

How do I find the best DSH plugins?

Start with our Editor's Picks — 13 hand-tested plugins our team actually uses. Then explore the full directory of 1837+ plugins organized by category, or read our in-depth reviews and comparison articles.

Is documentation available in Chinese?

Yes — full documentation in English at /docs/en/ and Simplified Chinese at /docs/zh/.

How is Oh! dsh different from the official plugin market?

The official plugin market is an install directory. Oh! dsh is a decision guide — we provide in-depth reviews, scenario-based recommendations, comparison articles, and best practices to help you choose the right plugins for your specific use case.

DeepSeek Harness turns Claude Code and Codex into sub-agents — and claims the scheduling layer

August 22, 2026 • 418 words • 2 min read

The most consequential thing DeepSeek Harness did this week wasn't a feature. It was a positioning move: Harness can now run Claude Code and Codex as sub-agents. That single capability re-frames what DSH is — not a competitor to those tools, but the layer above them.

What "sub-agent" means here

In DSH's model, "everything is a plugin" extends all the way to other agents. The same sub-agent machinery you'd use to spawn a fresh DSH child can now hand a task to a Claude Code or Codex backend instead. You write the orchestration once; the work runs wherever it runs best.

There's even community tooling pushing the idea further — unified sub-agent delegation across native spawn/fork plus Claude Code, Codex, Grok, and ACP bridges.

The "scheduling layer" thesis

This is the strategic bet, and it's worth naming plainly: whoever schedules the agents owns the layer above the models. Models are commodities-in-motion; agents are where the work happens; but the coordination — deciding who does what, with what context, under what trust boundary — is the high-value seat.

DSH is explicitly going for that seat. It doesn't need to be better than Claude Code at coding or better than Codex at cloud sandboxes. It needs to be the place you compose them from.

Why this is smart

Reuse instead of rebuild. Claude Code's terminal coding and Codex's cloud sandbox are years of work. A scheduler inherits them for free.
Best tool for the job. A single workflow can route to the agent that fits — coding to Claude Code, batch repo edits to Codex, internal steps to native DSH children.
Model-agnostic by construction. Since the child agents bring their own models, DSH stays neutral about which model wins underneath.

The honest caveats

Orchestration isn't free. Every delegation adds a hop: context marshalling, credential handling, and the trust question of letting one agent drive another. And the "no kernel, everything is a plugin" design that makes this possible is also what critics call its complexity tax. A scheduler is only as good as its failure modes — if a sub-agent hangs or misbehaves, the orchestrator owns it.

The takeaway

Running Claude Code and Codex as sub-agents is the clearest signal yet of what DSH is building toward: not "another coding agent", but the scheduling layer of the agent era. If you've been comparing DSH to Claude Code feature-by-feature, you've been asking the wrong question. The interesting question is which one ends up inside the other.